1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/mds/mds_reint.c
5 * Lustre Metadata Server (mds) reintegration routines
7 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 # define EXPORT_SYMTAB
31 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/jbd.h>
35 #include <linux/ext3_fs.h>
36 #include <linux/obd_support.h>
37 #include <linux/obd_class.h>
38 #include <linux/obd.h>
39 #include <linux/lustre_lib.h>
40 #include <linux/lustre_idl.h>
41 #include <linux/lustre_mds.h>
42 #include <linux/lustre_dlm.h>
43 #include <linux/lustre_log.h>
44 #include <linux/lustre_fsfilt.h>
45 #include "mds_internal.h"
47 struct mds_logcancel_data {
48 struct lov_mds_md *mlcd_lmm;
52 struct llog_cookie mlcd_cookies[0];
55 static void mds_cancel_cookies_cb(struct obd_device *obd,
56 __u64 transno, void *cb_data,
59 struct mds_logcancel_data *mlcd = cb_data;
60 struct lov_stripe_md *lsm = NULL;
61 struct llog_ctxt *ctxt;
64 obd_transno_commit_cb(obd, transno, error);
66 CDEBUG(D_HA, "cancelling %d cookies\n",
67 (int)(mlcd->mlcd_cookielen / sizeof(*mlcd->mlcd_cookies)));
69 rc = obd_unpackmd(obd->u.mds.mds_lov_exp, &lsm, mlcd->mlcd_lmm,
70 mlcd->mlcd_eadatalen);
72 CERROR("bad LSM cancelling %d log cookies: rc %d\n",
73 (int)(mlcd->mlcd_cookielen/sizeof(*mlcd->mlcd_cookies)),
76 ///* XXX 0 normally, SENDNOW for debug */);
77 ctxt = llog_get_context(&obd->obd_llogs,
78 mlcd->mlcd_cookies[0].lgc_subsys + 1);
79 rc = llog_cancel(ctxt, mlcd->mlcd_cookielen /
80 sizeof(*mlcd->mlcd_cookies),
81 mlcd->mlcd_cookies, OBD_LLOG_FL_SENDNOW, lsm);
83 CERROR("error cancelling %d log cookies: rc %d\n",
84 (int)(mlcd->mlcd_cookielen /
85 sizeof(*mlcd->mlcd_cookies)), rc);
88 OBD_FREE(mlcd, mlcd->mlcd_size);
91 /* Assumes caller has already pushed us into the kernel context. */
92 int mds_finish_transno(struct mds_obd *mds, struct inode *inode, void *handle,
93 struct ptlrpc_request *req, int rc, __u32 op_data)
95 struct mds_export_data *med = &req->rq_export->exp_mds_data;
96 struct obd_device *obd = req->rq_export->exp_obd;
97 struct mds_client_data *mcd = med->med_mcd;
98 int err, log_pri = D_HA;
103 /* if the export has already been failed, we have no last_rcvd slot */
104 if (req->rq_export->exp_failed) {
105 CERROR("committing transaction for disconnected client\n");
107 GOTO(out_commit, rc);
114 if (handle == NULL) {
115 /* if we're starting our own xaction, use our own inode */
116 inode = mds->mds_rcvd_filp->f_dentry->d_inode;
117 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
118 if (IS_ERR(handle)) {
119 CERROR("fsfilt_start: %ld\n", PTR_ERR(handle));
120 RETURN(PTR_ERR(handle));
126 transno = req->rq_reqmsg->transno;
128 LASSERTF(transno == 0, "BUG 3934, t"LPU64" rc %d\n", transno, rc);
129 } else if (transno == 0) {
130 spin_lock(&mds->mds_transno_lock);
131 transno = ++mds->mds_last_transno;
132 spin_unlock(&mds->mds_transno_lock);
134 spin_lock(&mds->mds_transno_lock);
135 if (transno > mds->mds_last_transno)
136 mds->mds_last_transno = transno;
137 spin_unlock(&mds->mds_transno_lock);
139 req->rq_repmsg->transno = req->rq_transno = transno;
140 mcd->mcd_last_transno = cpu_to_le64(transno);
141 mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
142 mcd->mcd_last_result = cpu_to_le32(rc);
143 mcd->mcd_last_data = cpu_to_le32(op_data);
145 fsfilt_add_journal_cb(obd, mds->mds_sb, transno, handle,
146 mds_commit_last_transno_cb, NULL);
148 err = fsfilt_write_record(obd, mds->mds_rcvd_filp, mcd,
149 sizeof(*mcd), &off, 0);
157 DEBUG_REQ(log_pri, req,
158 "wrote trans #"LPU64" client %s at idx %u: err = %d",
159 transno, mcd->mcd_uuid, med->med_idx, err);
161 err = mds_update_last_fid(obd, handle, 0);
168 err = mds_lov_write_objids(obd);
174 CDEBUG(log_pri, "wrote objids: err = %d\n", err);
178 err = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
180 CERROR("error committing transaction: %d\n", err);
188 /* this gives the same functionality as the code between
189 * sys_chmod and inode_setattr
190 * chown_common and inode_setattr
191 * utimes and inode_setattr
193 int mds_fix_attr(struct inode *inode, struct mds_update_record *rec)
195 time_t now = LTIME_S(CURRENT_TIME);
196 struct iattr *attr = &rec->ur_iattr;
197 unsigned int ia_valid = attr->ia_valid;
201 /* only fix up attrs if the client VFS didn't already */
202 if (!(ia_valid & ATTR_RAW))
205 if (!(ia_valid & ATTR_CTIME_SET))
206 LTIME_S(attr->ia_ctime) = now;
207 if (!(ia_valid & ATTR_ATIME_SET))
208 LTIME_S(attr->ia_atime) = now;
209 if (!(ia_valid & ATTR_MTIME_SET))
210 LTIME_S(attr->ia_mtime) = now;
212 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
216 if ((ia_valid & (ATTR_MTIME|ATTR_ATIME)) == (ATTR_MTIME|ATTR_ATIME)) {
217 if (rec->ur_fsuid != inode->i_uid &&
218 (error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
222 if (ia_valid & ATTR_SIZE) {
223 if ((error = ll_permission(inode, MAY_WRITE, NULL)) != 0)
227 if (ia_valid & ATTR_UID) {
230 if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
232 if (attr->ia_uid == (uid_t) -1)
233 attr->ia_uid = inode->i_uid;
234 if (attr->ia_gid == (gid_t) -1)
235 attr->ia_gid = inode->i_gid;
236 attr->ia_mode = inode->i_mode;
238 * If the user or group of a non-directory has been
239 * changed by a non-root user, remove the setuid bit.
240 * 19981026 David C Niemi <niemi@tux.org>
242 * Changed this to apply to all users, including root,
243 * to avoid some races. This is the behavior we had in
244 * 2.0. The check for non-root was definitely wrong
245 * for 2.2 anyway, as it should have been using
246 * CAP_FSETID rather than fsuid -- 19990830 SD.
248 if ((inode->i_mode & S_ISUID) == S_ISUID &&
249 !S_ISDIR(inode->i_mode)) {
250 attr->ia_mode &= ~S_ISUID;
251 attr->ia_valid |= ATTR_MODE;
254 * Likewise, if the user or group of a non-directory
255 * has been changed by a non-root user, remove the
256 * setgid bit UNLESS there is no group execute bit
257 * (this would be a file marked for mandatory
258 * locking). 19981026 David C Niemi <niemi@tux.org>
260 * Removed the fsuid check (see the comment above) --
263 if (((inode->i_mode & (S_ISGID | S_IXGRP)) ==
264 (S_ISGID | S_IXGRP)) && !S_ISDIR(inode->i_mode)) {
265 attr->ia_mode &= ~S_ISGID;
266 attr->ia_valid |= ATTR_MODE;
268 } else if (ia_valid & ATTR_MODE) {
269 int mode = attr->ia_mode;
271 if (attr->ia_mode == (mode_t) -1)
272 attr->ia_mode = inode->i_mode;
274 (mode & S_IALLUGO) | (inode->i_mode & ~S_IALLUGO);
279 void mds_steal_ack_locks(struct ptlrpc_request *req)
281 struct obd_export *exp = req->rq_export;
282 struct list_head *tmp;
283 struct ptlrpc_reply_state *oldrep;
284 struct ptlrpc_service *svc;
285 struct llog_create_locks *lcl;
287 char str[PTL_NALFMT_SIZE];
290 /* CAVEAT EMPTOR: spinlock order */
291 spin_lock_irqsave (&exp->exp_lock, flags);
292 list_for_each (tmp, &exp->exp_outstanding_replies) {
293 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
295 if (oldrep->rs_xid != req->rq_xid)
298 if (oldrep->rs_msg.opc != req->rq_reqmsg->opc)
299 CERROR ("Resent req xid "LPX64" has mismatched opc: "
300 "new %d old %d\n", req->rq_xid,
301 req->rq_reqmsg->opc, oldrep->rs_msg.opc);
303 svc = oldrep->rs_srv_ni->sni_service;
304 spin_lock (&svc->srv_lock);
306 list_del_init (&oldrep->rs_exp_list);
308 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
310 oldrep->rs_nlocks, oldrep,
311 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_msg.opc,
312 ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
314 for (i = 0; i < oldrep->rs_nlocks; i++)
315 ptlrpc_save_lock(req,
316 &oldrep->rs_locks[i],
317 oldrep->rs_modes[i]);
318 oldrep->rs_nlocks = 0;
320 lcl = oldrep->rs_llog_locks;
321 oldrep->rs_llog_locks = NULL;
323 ptlrpc_save_llog_lock(req, lcl);
325 DEBUG_REQ(D_HA, req, "stole locks for");
326 ptlrpc_schedule_difficult_reply (oldrep);
328 spin_unlock (&svc->srv_lock);
329 spin_unlock_irqrestore (&exp->exp_lock, flags);
332 spin_unlock_irqrestore (&exp->exp_lock, flags);
335 void mds_req_from_mcd(struct ptlrpc_request *req, struct mds_client_data *mcd)
337 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
338 mcd->mcd_last_transno, mcd->mcd_last_result);
339 req->rq_repmsg->transno = req->rq_transno = mcd->mcd_last_transno;
340 req->rq_repmsg->status = req->rq_status = mcd->mcd_last_result;
342 mds_steal_ack_locks(req);
345 static void reconstruct_reint_setattr(struct mds_update_record *rec,
346 int offset, struct ptlrpc_request *req)
348 struct mds_export_data *med = &req->rq_export->exp_mds_data;
349 struct mds_body *body;
352 mds_req_from_mcd(req, med->med_mcd);
354 de = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
356 LASSERT(PTR_ERR(de) == req->rq_status);
360 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
361 mds_pack_inode2body(req2obd(req), body, de->d_inode, 0);
363 /* Don't return OST-specific attributes if we didn't just set them */
364 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
365 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
366 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
367 body->valid |= OBD_MD_FLMTIME;
368 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
369 body->valid |= OBD_MD_FLATIME;
374 /* In the raw-setattr case, we lock the child inode.
375 * In the write-back case or if being called from open, the client holds a lock
378 * We use the ATTR_FROM_OPEN flag to tell these cases apart. */
379 static int mds_reint_setattr(struct mds_update_record *rec, int offset,
380 struct ptlrpc_request *req, struct lustre_handle *lh)
382 struct mds_obd *mds = mds_req2mds(req);
383 struct obd_device *obd = req->rq_export->exp_obd;
384 struct mds_body *body;
386 struct inode *inode = NULL;
387 struct lustre_handle lockh[2] = {{0}, {0}};
390 struct mds_logcancel_data *mlcd = NULL;
391 int rc = 0, cleanup_phase = 0, err;
395 LASSERT(offset == 1);
397 DEBUG_REQ(D_INODE, req, "setattr "LPU64"/%u %x",
398 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
399 rec->ur_iattr.ia_valid);
401 MDS_CHECK_RESENT(req, reconstruct_reint_setattr(rec, offset, req));
402 MD_COUNTER_INCREMENT(obd, setattr);
404 if (rec->ur_iattr.ia_valid & ATTR_FROM_OPEN) {
405 de = mds_id2dentry(obd, rec->ur_id1, NULL);
407 GOTO(cleanup, rc = PTR_ERR(de));
409 __u64 lockpart = MDS_INODELOCK_UPDATE;
410 if (rec->ur_iattr.ia_valid & (ATTR_MODE | ATTR_UID | ATTR_GID))
411 lockpart |= MDS_INODELOCK_LOOKUP;
412 de = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
413 lockh, &parent_mode, NULL, 0, lockpart);
415 GOTO(cleanup, rc = PTR_ERR(de));
423 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
424 rec->ur_eadata != NULL)
427 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_SETATTR_WRITE, inode->i_sb);
429 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
431 GOTO(cleanup, rc = PTR_ERR(handle));
433 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
434 CDEBUG(D_INODE, "setting mtime %lu, ctime %lu\n",
435 LTIME_S(rec->ur_iattr.ia_mtime),
436 LTIME_S(rec->ur_iattr.ia_ctime));
437 rc = mds_fix_attr(inode, rec);
441 if (rec->ur_iattr.ia_valid & ATTR_ATTR_FLAG) /* ioctl */
442 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_SETFLAGS,
443 (long)&rec->ur_iattr.ia_attr_flags);
445 rc = fsfilt_setattr(obd, de, handle, &rec->ur_iattr, 0);
447 if (rc == 0 && (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
448 rec->ur_eadata != NULL) {
449 struct lov_stripe_md *lsm = NULL;
451 rc = ll_permission(inode, MAY_WRITE, NULL);
455 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, mds->mds_lov_exp,
456 0, &lsm, rec->ur_eadata);
460 obd_free_memmd(mds->mds_lov_exp, &lsm);
462 rc = fsfilt_set_md(obd, inode, handle, rec->ur_eadata,
468 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
469 mds_pack_inode2body(obd, body, inode, 0);
470 body->id1 = *rec->ur_id1;
472 /* Don't return OST-specific attributes if we didn't just set them */
473 if (rec->ur_iattr.ia_valid & ATTR_SIZE)
474 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
475 if (rec->ur_iattr.ia_valid & (ATTR_MTIME | ATTR_MTIME_SET))
476 body->valid |= OBD_MD_FLMTIME;
477 if (rec->ur_iattr.ia_valid & (ATTR_ATIME | ATTR_ATIME_SET))
478 body->valid |= OBD_MD_FLATIME;
480 if (rc == 0 && rec->ur_cookielen && !IS_ERR(mds->mds_lov_obd)) {
481 OBD_ALLOC(mlcd, sizeof(*mlcd) + rec->ur_cookielen +
484 mlcd->mlcd_size = sizeof(*mlcd) + rec->ur_cookielen +
486 mlcd->mlcd_eadatalen = rec->ur_eadatalen;
487 mlcd->mlcd_cookielen = rec->ur_cookielen;
488 mlcd->mlcd_lmm = (void *)&mlcd->mlcd_cookies +
489 mlcd->mlcd_cookielen;
490 memcpy(&mlcd->mlcd_cookies, rec->ur_logcookies,
491 mlcd->mlcd_cookielen);
492 memcpy(mlcd->mlcd_lmm, rec->ur_eadata,
493 mlcd->mlcd_eadatalen);
495 CERROR("unable to allocate log cancel data\n");
501 fsfilt_add_journal_cb(req->rq_export->exp_obd, mds->mds_sb, 0,
502 handle, mds_cancel_cookies_cb, mlcd);
503 err = mds_finish_transno(mds, inode, handle, req, rc, 0);
504 switch (cleanup_phase) {
506 if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) &&
507 rec->ur_eadata != NULL)
512 if (lockh[1].cookie != 0)
513 ldlm_lock_decref(lockh + 1, parent_mode);
516 ldlm_lock_decref(lockh, LCK_PW);
518 ptlrpc_save_lock (req, lockh, LCK_PW);
533 static void reconstruct_reint_create(struct mds_update_record *rec, int offset,
534 struct ptlrpc_request *req)
536 struct mds_export_data *med = &req->rq_export->exp_mds_data;
537 struct dentry *parent, *child;
538 struct mds_body *body;
541 mds_req_from_mcd(req, med->med_mcd);
543 if (req->rq_status) {
548 parent = mds_id2dentry(req2obd(req), rec->ur_id1, NULL);
549 LASSERT(!IS_ERR(parent));
550 child = ll_lookup_one_len(rec->ur_name, parent,
551 rec->ur_namelen - 1);
552 LASSERT(!IS_ERR(child));
553 if ((child->d_flags & DCACHE_CROSS_REF)) {
554 LASSERTF(child->d_inode == NULL, "BUG 3869\n");
555 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
556 mds_pack_dentry2body(req2obd(req), body, child, 1);
557 } else if (child->d_inode == NULL) {
558 DEBUG_REQ(D_ERROR, req, "parent "DLID4" name %s mode %o",
559 OLID4(rec->ur_id1), rec->ur_name, rec->ur_mode);
560 LASSERTF(child->d_inode != NULL, "BUG 3869\n");
562 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
563 mds_pack_inode2body(req2obd(req), body, child->d_inode, 1);
570 static int mds_reint_create(struct mds_update_record *rec, int offset,
571 struct ptlrpc_request *req,
572 struct lustre_handle *lh)
574 struct dentry *dparent = NULL;
575 struct mds_obd *mds = mds_req2mds(req);
576 struct obd_device *obd = req->rq_export->exp_obd;
577 struct dentry *dchild = NULL;
578 struct inode *dir = NULL;
580 struct lustre_handle lockh[2] = {{0}, {0}};
582 int rc = 0, err, type = rec->ur_mode & S_IFMT, cleanup_phase = 0;
584 struct dentry_params dp;
585 struct mea *mea = NULL;
589 LASSERT(offset == 1);
591 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name,
594 DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %s mode %o",
595 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
596 rec->ur_name, rec->ur_mode);
598 MDS_CHECK_RESENT(req, reconstruct_reint_create(rec, offset, req));
600 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
601 GOTO(cleanup, rc = -ESTALE);
603 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, LCK_PW,
604 lockh, &parent_mode, rec->ur_name,
605 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
606 if (IS_ERR(dparent)) {
607 rc = PTR_ERR(dparent);
608 CERROR("parent lookup error %d\n", rc);
611 cleanup_phase = 1; /* locked parent dentry */
612 dir = dparent->d_inode;
615 ldlm_lock_dump_handle(D_OTHER, lockh);
617 /* try to retrieve MEA data for this dir */
618 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
624 * dir is already splitted, check is requested filename should
625 * live at this MDS or at another one.
627 int i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
628 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
629 CDEBUG(D_OTHER, "inapropriate MDS(%d) for %lu/%u:%s."
630 " should be %lu(%d)\n",
631 mea->mea_master, dparent->d_inode->i_ino,
632 dparent->d_inode->i_generation, rec->ur_name,
633 (unsigned long)id_group(&mea->mea_ids[i]), i);
634 GOTO(cleanup, rc = -ERESTART);
638 dchild = ll_lookup_one_len(rec->ur_name, dparent,
639 rec->ur_namelen - 1);
640 if (IS_ERR(dchild)) {
641 rc = PTR_ERR(dchild);
642 CERROR("Can't find "DLID4"/%s, error %d\n",
643 OLID4(rec->ur_id1), rec->ur_name, rc);
647 cleanup_phase = 2; /* child dentry */
649 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb);
651 if (type == S_IFREG || type == S_IFDIR) {
652 rc = mds_try_to_split_dir(obd, dparent, &mea, 0, parent_mode);
653 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d/%d\n",
654 obd->obd_name, dparent->d_inode->i_ino,
655 dparent->d_inode->i_generation, rc, parent_mode);
657 /* dir got splitted */
658 GOTO(cleanup, rc = -ERESTART);
660 /* error happened during spitting. */
665 if (dir->i_mode & S_ISGID) {
666 if (S_ISDIR(rec->ur_mode))
667 rec->ur_mode |= S_ISGID;
671 * here inode number should be used only in the case of replaying. It is
672 * needed to check if object already created in the case of creating
675 dchild->d_fsdata = (void *)&dp;
676 dp.p_inum = (unsigned long)id_ino(rec->ur_id2);
681 handle = fsfilt_start(obd, dir, FSFILT_OP_CREATE, NULL);
683 GOTO(cleanup, rc = PTR_ERR(handle));
684 rc = ll_vfs_create(dir, dchild, rec->ur_mode, NULL);
690 struct lustre_id sid;
693 * as Peter asked, mkdir() should distribute new directories
694 * over the whole cluster in order to distribute namespace
695 * processing load. first, we calculate which MDS to use to put
696 * new directory's inode in.
698 i = mds_choose_mdsnum(obd, rec->ur_name, rec->ur_namelen - 1,
700 if (i == mds->mds_num) {
701 /* inode will be created locally */
702 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
704 GOTO(cleanup, rc = PTR_ERR(handle));
706 rc = vfs_mkdir(dir, dchild, rec->ur_mode);
708 CERROR("Can't create dir %s, rc = %d\n",
709 dchild->d_name.name, rc);
713 down(&dchild->d_inode->i_sem);
715 rc = mds_update_inode_sid(obd, dchild->d_inode,
716 handle, rec->ur_id2);
718 CERROR("mds_update_inode_sid() failed, inode %lu, "
719 "rc %d\n", dchild->d_inode->i_ino, rc);
723 * make sure, that fid is up-to-date.
725 mds_set_last_fid(obd, id_fid(rec->ur_id2));
727 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
730 CERROR("mds_alloc_inode_sid() failed, inode %lu, "
731 "rc %d\n", dchild->d_inode->i_ino, rc);
734 up(&dchild->d_inode->i_sem);
740 nstripes = *(u16 *)rec->ur_eadata;
742 if (rc == 0 && nstripes) {
744 * we pass LCK_EX to split routine to signal,
745 * that we have exclusive access to the
746 * directory. Simple because nobody knows it
747 * already exists -bzzz
749 rc = mds_try_to_split_dir(obd, dchild,
753 /* dir got splitted */
756 /* an error occured during
761 } else if (!DENTRY_VALID(dchild)) {
762 /* inode will be created on another MDS */
763 struct obdo *oa = NULL;
764 struct mds_body *body;
766 /* first, create that inode */
769 GOTO(cleanup, rc = -ENOMEM);
774 if (rec->ur_eadata) {
775 /* user asks for creating splitted dir */
776 oa->o_easize = *((u16 *) rec->ur_eadata);
779 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
780 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
781 OBD_MD_FLUID | OBD_MD_FLGID);
783 oa->o_mode = dir->i_mode;
785 CDEBUG(D_OTHER, "%s: create dir on MDS %u\n",
788 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
790 * here inode number and generation are
791 * important, as this is replay request and we
792 * need them to check if such an object is
795 CDEBUG(D_HA, "%s: replay dir creation %*s -> %u/%u\n",
796 obd->obd_name, rec->ur_namelen - 1,
797 rec->ur_name, (unsigned)id_ino(rec->ur_id2),
798 (unsigned)id_gen(rec->ur_id2));
799 oa->o_id = id_ino(rec->ur_id2);
800 oa->o_fid = id_fid(rec->ur_id2);
801 oa->o_generation = id_gen(rec->ur_id2);
802 oa->o_flags |= OBD_FL_RECREATE_OBJS;
805 /* before obd_create() is called, o_fid is not known. */
806 rc = obd_create(mds->mds_lmv_exp, oa, NULL, NULL);
808 CERROR("can't create remote inode: %d\n", rc);
809 DEBUG_REQ(D_ERROR, req, "parent "LPU64"/%u name %s mode %o",
810 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
811 rec->ur_name, rec->ur_mode);
816 /* now, add new dir entry for it */
817 handle = fsfilt_start(obd, dir, FSFILT_OP_MKDIR, NULL);
818 if (IS_ERR(handle)) {
820 GOTO(cleanup, rc = PTR_ERR(handle));
823 /* creating local dentry for remote inode. */
824 rc = fsfilt_add_dir_entry(obd, dparent, rec->ur_name,
825 rec->ur_namelen - 1, oa->o_id,
826 oa->o_generation, i, oa->o_fid);
829 CERROR("Can't create local entry %*s for "
830 "remote inode.\n", rec->ur_namelen - 1,
836 body = lustre_msg_buf(req->rq_repmsg,
838 body->valid |= OBD_MD_FLID | OBD_MD_MDS |
841 obdo2id(&body->id1, oa);
844 /* requested name exists in the directory */
851 handle = fsfilt_start(obd, dir, FSFILT_OP_SYMLINK, NULL);
853 GOTO(cleanup, rc = PTR_ERR(handle));
854 if (rec->ur_tgt == NULL) /* no target supplied */
855 rc = -EINVAL; /* -EPROTO? */
857 rc = ll_vfs_symlink(dir, dchild, rec->ur_tgt, S_IALLUGO);
865 int rdev = rec->ur_rdev;
866 handle = fsfilt_start(obd, dir, FSFILT_OP_MKNOD, NULL);
868 GOTO(cleanup, (handle = NULL, rc = PTR_ERR(handle)));
869 rc = vfs_mknod(dir, dchild, rec->ur_mode, rdev);
874 CERROR("bad file type %o creating %s\n", type, rec->ur_name);
875 dchild->d_fsdata = NULL;
876 GOTO(cleanup, rc = -EINVAL);
879 /* In case we stored the desired inum in here, we want to clean up. */
880 if (dchild->d_fsdata == (void *)(unsigned long)id_ino(rec->ur_id2))
881 dchild->d_fsdata = NULL;
884 CDEBUG(D_INODE, "error during create: %d\n", rc);
886 } else if (dchild->d_inode) {
888 struct mds_body *body;
889 struct inode *inode = dchild->d_inode;
892 iattr.ia_uid = rec->ur_fsuid;
893 LTIME_S(iattr.ia_atime) = rec->ur_time;
894 LTIME_S(iattr.ia_ctime) = rec->ur_time;
895 LTIME_S(iattr.ia_mtime) = rec->ur_time;
897 if (dir->i_mode & S_ISGID)
898 iattr.ia_gid = dir->i_gid;
900 iattr.ia_gid = rec->ur_fsgid;
902 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
903 ATTR_MTIME | ATTR_CTIME;
905 if (id_ino(rec->ur_id2)) {
906 LASSERT(id_ino(rec->ur_id2) == inode->i_ino);
907 inode->i_generation = id_gen(rec->ur_id2);
909 if (type != S_IFDIR) {
911 * updating inode self id, as inode already
912 * exists and we should make sure, its sid will
913 * be the same as we reveived.
916 rc = mds_update_inode_sid(obd, inode,
917 handle, rec->ur_id2);
920 CERROR("Can't update inode self id, "
925 * make sure, that fid is up-to-date.
927 mds_set_last_fid(obd, id_fid(rec->ur_id2));
930 /* dirtied and committed by the upcoming setattr. */
931 CDEBUG(D_INODE, "recreated ino %lu with gen %u\n",
932 inode->i_ino, inode->i_generation);
934 struct lustre_handle child_ino_lockh;
936 CDEBUG(D_INODE, "created ino %lu with gen %x\n",
937 inode->i_ino, inode->i_generation);
939 if (type != S_IFDIR) {
940 struct lustre_id sid;
943 * allocate new id for @inode if it is not dir,
944 * because for dir it was already done.
947 rc = mds_alloc_inode_sid(obd, inode,
951 CERROR("mds_alloc_inode_sid() failed, "
952 "inode %lu, rc %d\n", inode->i_ino,
959 * the inode we were allocated may have just
960 * been freed by an unlink operation. We take
961 * this lock to synchronize against the matching
962 * reply-ack-lock taken in unlink, to avoid
963 * replay problems if this reply makes it out to
964 * the client but the unlink's does not. See
965 * bug 2029 for more detail.
967 rc = mds_lock_new_child(obd, inode, &child_ino_lockh);
968 if (rc != ELDLM_OK) {
969 CERROR("error locking for unlink/create sync: "
972 ldlm_lock_decref(&child_ino_lockh, LCK_EX);
977 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
979 CERROR("error on child setattr: rc = %d\n", rc);
981 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
982 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
984 CERROR("error on parent setattr: rc = %d\n", rc);
986 MD_COUNTER_INCREMENT(obd, create);
988 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
989 mds_pack_inode2body(obd, body, inode, 1);
994 err = mds_finish_transno(mds, dir, handle, req, rc, 0);
997 /* Destroy the file we just created. This should not need extra
998 * journal credits, as we have already modified all of the
999 * blocks needed in order to create the file in the first
1003 err = vfs_rmdir(dir, dchild);
1005 CERROR("rmdir in error path: %d\n", err);
1008 err = vfs_unlink(dir, dchild);
1010 CERROR("unlink in error path: %d\n", err);
1016 switch (cleanup_phase) {
1017 case 2: /* child dentry */
1019 case 1: /* locked parent dentry */
1021 if (lockh[1].cookie != 0)
1022 ldlm_lock_decref(lockh + 1, parent_mode);
1025 ldlm_lock_decref(lockh, LCK_PW);
1027 ptlrpc_save_lock(req, lockh, LCK_PW);
1033 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1037 OBD_FREE(mea, mea_size);
1038 req->rq_status = rc;
1043 res_gt(struct ldlm_res_id *res1, struct ldlm_res_id *res2,
1044 ldlm_policy_data_t *p1, ldlm_policy_data_t *p2)
1048 for (i = 0; i < RES_NAME_SIZE; i++) {
1050 * this is needed to make zeroed res_id entries to be put at the
1051 * end of list in *ordered_locks() .
1053 if (res1->name[i] == 0 && res2->name[i] != 0)
1055 if (res2->name[i] == 0 && res1->name[i] != 0)
1057 if (res1->name[i] > res2->name[i])
1059 if (res1->name[i] < res2->name[i])
1066 if (memcmp(p1, p2, sizeof(*p1)) < 0)
1072 /* This function doesn't use ldlm_match_or_enqueue because we're always called
1073 * with EX or PW locks, and the MDS is no longer allowed to match write locks,
1074 * because they take the place of local semaphores.
1076 * One or two locks are taken in numerical order. A res_id->name[0] of 0 means
1077 * no lock is taken for that res_id. Must be at least one non-zero res_id. */
1078 int enqueue_ordered_locks(struct obd_device *obd, struct ldlm_res_id *p1_res_id,
1079 struct lustre_handle *p1_lockh, int p1_lock_mode,
1080 ldlm_policy_data_t *p1_policy,
1081 struct ldlm_res_id *p2_res_id,
1082 struct lustre_handle *p2_lockh, int p2_lock_mode,
1083 ldlm_policy_data_t *p2_policy)
1085 int lock_modes[2] = { p1_lock_mode, p2_lock_mode };
1086 struct ldlm_res_id *res_id[2] = { p1_res_id, p2_res_id };
1087 struct lustre_handle *handles[2] = { p1_lockh, p2_lockh };
1088 ldlm_policy_data_t *policies[2] = { p1_policy, p2_policy };
1092 LASSERT(p1_res_id != NULL && p2_res_id != NULL);
1094 CDEBUG(D_INFO, "locks before: "LPU64"/"LPU64"\n",
1095 res_id[0]->name[0], res_id[1]->name[0]);
1097 if (res_gt(p1_res_id, p2_res_id, p1_policy, p2_policy)) {
1098 handles[1] = p1_lockh;
1099 handles[0] = p2_lockh;
1100 res_id[1] = p1_res_id;
1101 res_id[0] = p2_res_id;
1102 lock_modes[1] = p1_lock_mode;
1103 lock_modes[0] = p2_lock_mode;
1104 policies[1] = p1_policy;
1105 policies[0] = p2_policy;
1108 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"\n",
1109 res_id[0]->name[0], res_id[1]->name[0]);
1111 flags = LDLM_FL_LOCAL_ONLY;
1112 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, *res_id[0],
1113 LDLM_IBITS, policies[0], lock_modes[0], &flags,
1114 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
1115 NULL, 0, NULL, handles[0]);
1118 ldlm_lock_dump_handle(D_OTHER, handles[0]);
1120 if (!memcmp(res_id[0], res_id[1], sizeof(*res_id[0])) &&
1121 (policies[0]->l_inodebits.bits & policies[1]->l_inodebits.bits)) {
1122 memcpy(handles[1], handles[0], sizeof(*(handles[1])));
1123 ldlm_lock_addref(handles[1], lock_modes[1]);
1124 } else if (res_id[1]->name[0] != 0) {
1125 flags = LDLM_FL_LOCAL_ONLY;
1126 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1127 *res_id[1], LDLM_IBITS, policies[1],
1128 lock_modes[1], &flags, mds_blocking_ast,
1129 ldlm_completion_ast, NULL, NULL, NULL, 0,
1131 if (rc != ELDLM_OK) {
1132 ldlm_lock_decref(handles[0], lock_modes[0]);
1135 ldlm_lock_dump_handle(D_OTHER, handles[1]);
1141 int enqueue_4ordered_locks(struct obd_device *obd,struct ldlm_res_id *p1_res_id,
1142 struct lustre_handle *p1_lockh, int p1_lock_mode,
1143 ldlm_policy_data_t *p1_policy,
1144 struct ldlm_res_id *p2_res_id,
1145 struct lustre_handle *p2_lockh, int p2_lock_mode,
1146 ldlm_policy_data_t *p2_policy,
1147 struct ldlm_res_id *c1_res_id,
1148 struct lustre_handle *c1_lockh, int c1_lock_mode,
1149 ldlm_policy_data_t *c1_policy,
1150 struct ldlm_res_id *c2_res_id,
1151 struct lustre_handle *c2_lockh, int c2_lock_mode,
1152 ldlm_policy_data_t *c2_policy)
1154 struct ldlm_res_id *res_id[5] = { p1_res_id, p2_res_id,
1155 c1_res_id, c2_res_id };
1156 struct lustre_handle *dlm_handles[5] = { p1_lockh, p2_lockh,
1157 c1_lockh, c2_lockh };
1158 int lock_modes[5] = { p1_lock_mode, p2_lock_mode,
1159 c1_lock_mode, c2_lock_mode };
1160 ldlm_policy_data_t *policies[5] = { p1_policy, p2_policy,
1161 c1_policy, c2_policy};
1162 int rc, i, j, sorted, flags;
1165 CDEBUG(D_DLMTRACE, "locks before: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1166 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1167 res_id[3]->name[0]);
1170 * simple insertion sort - we have at most 4 elements. Note, that zeroed
1171 * res_id should be at the end of list after sorting is finished.
1173 for (i = 1; i < 4; i++) {
1175 dlm_handles[4] = dlm_handles[i];
1176 res_id[4] = res_id[i];
1177 lock_modes[4] = lock_modes[i];
1178 policies[4] = policies[i];
1182 if (res_gt(res_id[j], res_id[4], policies[j],
1184 dlm_handles[j + 1] = dlm_handles[j];
1185 res_id[j + 1] = res_id[j];
1186 lock_modes[j + 1] = lock_modes[j];
1187 policies[j + 1] = policies[j];
1192 } while (j >= 0 && !sorted);
1194 dlm_handles[j + 1] = dlm_handles[4];
1195 res_id[j + 1] = res_id[4];
1196 lock_modes[j + 1] = lock_modes[4];
1197 policies[j + 1] = policies[4];
1200 CDEBUG(D_DLMTRACE, "lock order: "LPU64"/"LPU64"/"LPU64"/"LPU64"\n",
1201 res_id[0]->name[0], res_id[1]->name[0], res_id[2]->name[0],
1202 res_id[3]->name[0]);
1204 /* XXX we could send ASTs on all these locks first before blocking? */
1205 for (i = 0; i < 4; i++) {
1209 * nevertheless zeroed res_ids should be at the end of list, and
1210 * could use break here, I think, that it is more correctly for
1211 * clear understanding of code to have continue here, as it
1212 * clearly means, that zeroed res_id should be skipped and does
1213 * not mean, that if we meet zeroed res_id we should stop
1216 if (res_id[i]->name[0] == 0)
1220 !memcmp(res_id[i], res_id[i-1], sizeof(*res_id[i])) &&
1221 (policies[i]->l_inodebits.bits &
1222 policies[i-1]->l_inodebits.bits) ) {
1223 memcpy(dlm_handles[i], dlm_handles[i-1],
1224 sizeof(*(dlm_handles[i])));
1225 ldlm_lock_addref(dlm_handles[i], lock_modes[i]);
1227 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1228 *res_id[i], LDLM_IBITS,
1230 lock_modes[i], &flags,
1232 ldlm_completion_ast, NULL, NULL,
1233 NULL, 0, NULL, dlm_handles[i]);
1235 GOTO(out_err, rc = -EIO);
1236 ldlm_lock_dump_handle(D_OTHER, dlm_handles[i]);
1243 ldlm_lock_decref(dlm_handles[i], lock_modes[i]);
1248 /* In the unlikely case that the child changed while we were waiting
1249 * on the lock, we need to drop the lock on the old child and either:
1250 * - if the child has a lower resource name, then we have to also
1251 * drop the parent lock and regain the locks in the right order
1252 * - in the rename case, if the child has a lower resource name than one of
1253 * the other parent/child resources (maxres) we also need to reget the locks
1254 * - if the child has a higher resource name (this is the common case)
1255 * we can just get the lock on the new child (still in lock order)
1257 * Returns 0 if the child did not change or if it changed but could be locked.
1258 * Returns 1 if the child changed and we need to re-lock (no locks held).
1259 * Returns -ve error with a valid dchild (no locks held). */
1260 static int mds_verify_child(struct obd_device *obd,
1261 struct ldlm_res_id *parent_res_id,
1262 struct lustre_handle *parent_lockh,
1263 struct dentry *dparent, int parent_mode,
1264 struct ldlm_res_id *child_res_id,
1265 struct lustre_handle *child_lockh,
1266 struct dentry **dchildp, int child_mode,
1267 ldlm_policy_data_t *child_policy,
1268 const char *name, int namelen,
1269 struct ldlm_res_id *maxres,
1270 unsigned long child_ino,
1273 struct lustre_id sid;
1274 struct dentry *vchild, *dchild = *dchildp;
1275 int rc = 0, cleanup_phase = 2; /* parent, child locks */
1278 vchild = ll_lookup_one_len(name, dparent, namelen - 1);
1280 GOTO(cleanup, rc = PTR_ERR(vchild));
1282 if ((vchild->d_flags & DCACHE_CROSS_REF)) {
1283 if (child_gen == vchild->d_generation &&
1284 child_ino == vchild->d_inum) {
1293 if (likely((vchild->d_inode == NULL && child_res_id->name[0] == 0) ||
1294 (vchild->d_inode != NULL &&
1295 child_gen == vchild->d_inode->i_generation &&
1296 child_ino == vchild->d_inode->i_ino))) {
1304 CDEBUG(D_DLMTRACE, "child inode changed: %p != %p (%lu != "LPU64")\n",
1305 vchild->d_inode, dchild ? dchild->d_inode : 0,
1306 vchild->d_inode ? vchild->d_inode->i_ino : 0,
1307 child_res_id->name[0]);
1309 if (child_res_id->name[0] != 0)
1310 ldlm_lock_decref(child_lockh, child_mode);
1314 cleanup_phase = 1; /* parent lock only */
1315 *dchildp = dchild = vchild;
1317 if (dchild->d_inode || (dchild->d_flags & DCACHE_CROSS_REF)) {
1320 if (dchild->d_inode) {
1321 down(&dchild->d_inode->i_sem);
1322 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1323 up(&dchild->d_inode->i_sem);
1325 CERROR("Can't read inode self id, inode %lu,"
1326 " rc %d\n", dchild->d_inode->i_ino, rc);
1329 child_res_id->name[0] = id_fid(&sid);
1330 child_res_id->name[1] = id_group(&sid);
1332 child_res_id->name[0] = dchild->d_fid;
1333 child_res_id->name[1] = dchild->d_mdsnum;
1336 if (res_gt(parent_res_id, child_res_id, NULL, NULL) ||
1337 res_gt(maxres, child_res_id, NULL, NULL)) {
1338 CDEBUG(D_DLMTRACE, "relock "LPU64"<("LPU64"|"LPU64")\n",
1339 child_res_id->name[0], parent_res_id->name[0],
1341 GOTO(cleanup, rc = 1);
1344 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1345 *child_res_id, LDLM_IBITS, child_policy,
1346 child_mode, &flags, mds_blocking_ast,
1347 ldlm_completion_ast, NULL, NULL, NULL, 0,
1350 GOTO(cleanup, rc = -EIO);
1353 memset(child_res_id, 0, sizeof(*child_res_id));
1359 switch(cleanup_phase) {
1361 if (child_res_id->name[0] != 0)
1362 ldlm_lock_decref(child_lockh, child_mode);
1364 ldlm_lock_decref(parent_lockh, parent_mode);
1370 int mds_get_parent_child_locked(struct obd_device *obd, struct mds_obd *mds,
1371 struct lustre_id *id,
1372 struct lustre_handle *parent_lockh,
1373 struct dentry **dparentp, int parent_mode,
1374 __u64 parent_lockpart, int *update_mode,
1375 char *name, int namelen,
1376 struct lustre_handle *child_lockh,
1377 struct dentry **dchildp, int child_mode,
1378 __u64 child_lockpart)
1380 ldlm_policy_data_t parent_policy = {.l_inodebits = { parent_lockpart }};
1381 ldlm_policy_data_t child_policy = {.l_inodebits = { child_lockpart }};
1382 struct ldlm_res_id parent_res_id = { .name = {0} };
1383 struct ldlm_res_id child_res_id = { .name = {0} };
1384 int rc = 0, cleanup_phase = 0;
1385 unsigned long child_ino;
1386 struct lustre_id sid;
1387 __u32 child_gen = 0;
1388 struct inode *inode;
1391 /* Step 1: Lookup parent */
1392 *dparentp = mds_id2dentry(obd, id, NULL);
1393 if (IS_ERR(*dparentp)) {
1394 rc = PTR_ERR(*dparentp);
1399 CDEBUG(D_INODE, "parent ino %lu, name %s\n",
1400 (*dparentp)->d_inode->i_ino, name);
1402 parent_res_id.name[0] = id_fid(id);
1403 parent_res_id.name[1] = id_group(id);
1406 parent_lockh[1].cookie = 0;
1407 if (name && IS_PDIROPS((*dparentp)->d_inode)) {
1408 struct ldlm_res_id res_id = { .name = {0} };
1409 ldlm_policy_data_t policy;
1412 *update_mode = mds_lock_mode_for_dir(obd, *dparentp, parent_mode);
1414 res_id.name[0] = id_fid(id);
1415 res_id.name[1] = id_group(id);
1416 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1418 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1419 res_id, LDLM_IBITS, &policy,
1420 *update_mode, &flags,
1422 ldlm_completion_ast,
1423 NULL, NULL, NULL, 0, NULL,
1429 parent_res_id.name[2] = full_name_hash(name, namelen - 1);
1431 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
1432 (unsigned long)id_fid(id), (unsigned long)id_group(id),
1433 parent_res_id.name[2]);
1437 cleanup_phase = 1; /* parent dentry */
1439 /* Step 2: Lookup child (without DLM lock, to get resource name) */
1440 *dchildp = ll_lookup_one_len(name, *dparentp, namelen - 1);
1441 if (IS_ERR(*dchildp)) {
1442 rc = PTR_ERR(*dchildp);
1443 CDEBUG(D_INODE, "child lookup error %d\n", rc);
1447 if ((*dchildp)->d_flags & DCACHE_CROSS_REF) {
1449 * inode lives on another MDS: return * fid/mdsnum and LOOKUP
1450 * lock. Drop possible UPDATE lock!
1452 child_policy.l_inodebits.bits &= ~MDS_INODELOCK_UPDATE;
1453 child_policy.l_inodebits.bits |= MDS_INODELOCK_LOOKUP;
1455 child_res_id.name[0] = (*dchildp)->d_fid;
1456 child_res_id.name[1] = (*dchildp)->d_mdsnum;
1457 child_gen = (*dchildp)->d_generation;
1458 child_ino = (*dchildp)->d_inum;
1462 inode = (*dchildp)->d_inode;
1464 inode = igrab(inode);
1468 down(&inode->i_sem);
1469 rc = mds_read_inode_sid(obd, inode, &sid);
1472 CERROR("Can't read inode self id, inode %lu, "
1473 "rc %d\n", inode->i_ino, rc);
1478 child_res_id.name[0] = id_fid(&sid);
1479 child_res_id.name[1] = id_group(&sid);
1480 child_gen = inode->i_generation;
1481 child_ino = inode->i_ino;
1485 cleanup_phase = 2; /* child dentry */
1487 /* Step 3: Lock parent and child in resource order. If child doesn't
1488 * exist, we still have to lock the parent and re-lookup. */
1489 rc = enqueue_ordered_locks(obd, &parent_res_id, parent_lockh, parent_mode,
1490 &parent_policy, &child_res_id, child_lockh,
1491 child_mode, &child_policy);
1495 if ((*dchildp)->d_inode || ((*dchildp)->d_flags & DCACHE_CROSS_REF))
1496 cleanup_phase = 4; /* child lock */
1498 cleanup_phase = 3; /* parent lock */
1500 /* Step 4: Re-lookup child to verify it hasn't changed since locking */
1501 rc = mds_verify_child(obd, &parent_res_id, parent_lockh, *dparentp,
1502 parent_mode, &child_res_id, child_lockh,
1503 dchildp, child_mode, &child_policy,
1504 name, namelen, &parent_res_id,
1505 child_ino, child_gen);
1516 switch (cleanup_phase) {
1518 ldlm_lock_decref(child_lockh, child_mode);
1520 ldlm_lock_decref(parent_lockh, parent_mode);
1525 if (parent_lockh[1].cookie)
1526 ldlm_lock_decref(parent_lockh + 1, *update_mode);
1534 void mds_reconstruct_generic(struct ptlrpc_request *req)
1536 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1537 mds_req_from_mcd(req, med->med_mcd);
1540 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
1541 * we instead link the inode into the PENDING directory until it is
1542 * finally released. We can't simply call mds_reint_rename() or some
1543 * part thereof, because we don't have the inode to check for link
1544 * count/open status until after it is locked.
1546 * For lock ordering, caller must get child->i_sem first, then pending->i_sem
1547 * before starting journal transaction.
1549 * returns 1 on success
1550 * returns 0 if we lost a race and didn't make a new link
1551 * returns negative on error
1553 static int mds_orphan_add_link(struct mds_update_record *rec,
1554 struct obd_device *obd, struct dentry *dentry)
1556 struct mds_obd *mds = &obd->u.mds;
1557 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1558 struct inode *inode = dentry->d_inode;
1559 struct dentry *pending_child;
1560 char idname[LL_ID_NAMELEN];
1561 int idlen = 0, rc, mode;
1564 LASSERT(inode != NULL);
1565 LASSERT(!mds_inode_is_orphan(inode));
1566 #ifndef HAVE_I_ALLOC_SEM
1567 LASSERT(down_trylock(&inode->i_sem) != 0);
1569 LASSERT(down_trylock(&pending_dir->i_sem) != 0);
1571 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1573 CDEBUG(D_INODE, "pending destroy of %dx open %d linked %s %s = %s\n",
1574 mds_orphan_open_count(inode), inode->i_nlink,
1575 S_ISDIR(inode->i_mode) ? "dir" :
1576 S_ISREG(inode->i_mode) ? "file" : "other",
1577 rec->ur_name, idname);
1579 if (mds_orphan_open_count(inode) == 0 || inode->i_nlink != 0)
1582 pending_child = lookup_one_len(idname, mds->mds_pending_dir, idlen);
1583 if (IS_ERR(pending_child))
1584 RETURN(PTR_ERR(pending_child));
1586 if (pending_child->d_inode != NULL) {
1587 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
1588 LASSERT(pending_child->d_inode == inode);
1589 GOTO(out_dput, rc = 0);
1592 /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
1593 * for linking and return real mode back then -bzzz */
1594 mode = inode->i_mode;
1595 inode->i_mode = S_IFREG;
1596 rc = vfs_link(dentry, pending_dir, pending_child);
1598 CERROR("error linking orphan %s to PENDING: rc = %d\n",
1601 mds_inode_set_orphan(inode);
1603 /* return mode and correct i_nlink if inode is directory */
1604 inode->i_mode = mode;
1605 LASSERTF(inode->i_nlink == 1, "%s nlink == %d\n",
1606 S_ISDIR(mode) ? "dir" : S_ISREG(mode) ? "file" : "other",
1608 if (S_ISDIR(mode)) {
1610 pending_dir->i_nlink++;
1611 mark_inode_dirty(inode);
1612 mark_inode_dirty(pending_dir);
1617 l_dput(pending_child);
1621 int mds_create_local_dentry(struct mds_update_record *rec,
1622 struct obd_device *obd)
1624 struct mds_obd *mds = &obd->u.mds;
1625 struct inode *id_dir = mds->mds_id_dir->d_inode;
1626 int idlen = 0, rc, cleanup_phase = 0;
1627 struct dentry *new_child = NULL;
1628 char *idname = rec->ur_name;
1629 struct dentry *child = NULL;
1630 struct lustre_handle lockh[2] = {{0}, {0}};
1631 struct lustre_id sid;
1635 down(&id_dir->i_sem);
1636 idlen = ll_id2str(idname, id_ino(rec->ur_id1),
1637 id_gen(rec->ur_id1));
1639 CDEBUG(D_OTHER, "look for local dentry '%s' for "DLID4"\n",
1640 idname, OLID4(rec->ur_id1));
1642 new_child = ll_lookup_one_len(idname, mds->mds_id_dir,
1645 if (IS_ERR(new_child)) {
1646 CERROR("can't lookup %s: %d\n", idname,
1647 (int) PTR_ERR(new_child));
1648 GOTO(cleanup, rc = PTR_ERR(new_child));
1652 down(&id_dir->i_sem);
1653 rc = mds_read_inode_sid(obd, id_dir, &sid);
1656 CERROR("Can't read inode self id, inode %lu, "
1657 "rc %d\n", id_dir->i_ino, rc);
1661 if (new_child->d_inode != NULL) {
1662 /* nice. we've already have local dentry! */
1663 CDEBUG(D_OTHER, "found dentry in FIDS/: %u/%u\n",
1664 (unsigned)new_child->d_inode->i_ino,
1665 (unsigned)new_child->d_inode->i_generation);
1667 id_ino(rec->ur_id1) = id_dir->i_ino;
1668 id_gen(rec->ur_id1) = id_dir->i_generation;
1669 rec->ur_namelen = idlen + 1;
1671 id_fid(rec->ur_id1) = id_fid(&sid);
1672 id_group(rec->ur_id1) = id_group(&sid);
1674 GOTO(cleanup, rc = 0);
1677 /* new, local dentry will be added soon. we need no aliases here */
1680 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1681 child = mds_id2dentry(obd, rec->ur_id1, NULL);
1683 child = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1684 LCK_EX, lockh, NULL, NULL, 0,
1685 MDS_INODELOCK_UPDATE);
1688 if (IS_ERR(child)) {
1689 rc = PTR_ERR(child);
1690 if (rc != -ENOENT || !(rec->ur_mode & MDS_MODE_REPLAY))
1691 CERROR("can't get victim: %d\n", rc);
1696 handle = fsfilt_start(obd, id_dir, FSFILT_OP_LINK, NULL);
1698 GOTO(cleanup, rc = PTR_ERR(handle));
1700 rc = fsfilt_add_dir_entry(obd, mds->mds_id_dir, idname,
1701 idlen, id_ino(rec->ur_id1),
1702 id_gen(rec->ur_id1), mds->mds_num,
1703 id_fid(rec->ur_id1));
1705 CERROR("error linking orphan %lu/%lu to FIDS: rc = %d\n",
1706 (unsigned long)child->d_inode->i_ino,
1707 (unsigned long)child->d_inode->i_generation, rc);
1709 if (S_ISDIR(child->d_inode->i_mode)) {
1711 mark_inode_dirty(id_dir);
1713 mark_inode_dirty(child->d_inode);
1715 fsfilt_commit(obd, mds->mds_sb, id_dir, handle, 0);
1717 id_ino(rec->ur_id1) = id_dir->i_ino;
1718 id_gen(rec->ur_id1) = id_dir->i_generation;
1719 rec->ur_namelen = idlen + 1;
1721 id_fid(rec->ur_id1) = id_fid(&sid);
1722 id_group(rec->ur_id1) = id_group(&sid);
1726 switch(cleanup_phase) {
1728 if (!(rec->ur_mode & MDS_MODE_DONT_LOCK))
1729 ldlm_lock_decref(lockh, LCK_EX);
1739 static int mds_copy_unlink_reply(struct ptlrpc_request *master,
1740 struct ptlrpc_request *slave)
1742 void *cookie, *cookie2;
1743 struct mds_body *body2;
1744 struct mds_body *body;
1748 body = lustre_msg_buf(slave->rq_repmsg, 0, sizeof(*body));
1749 LASSERT(body != NULL);
1751 body2 = lustre_msg_buf(master->rq_repmsg, 0, sizeof (*body));
1752 LASSERT(body2 != NULL);
1754 if (!(body->valid & (OBD_MD_FLID | OBD_MD_FLGENER)))
1757 memcpy(body2, body, sizeof(*body));
1758 body2->valid &= ~OBD_MD_FLCOOKIE;
1760 if (!(body->valid & OBD_MD_FLEASIZE) &&
1761 !(body->valid & OBD_MD_FLDIREA))
1764 if (body->eadatasize == 0) {
1765 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
1769 LASSERT(master->rq_repmsg->buflens[1] >= body->eadatasize);
1771 ea = lustre_msg_buf(slave->rq_repmsg, 1, body->eadatasize);
1772 LASSERT(ea != NULL);
1774 ea2 = lustre_msg_buf(master->rq_repmsg, 1, body->eadatasize);
1775 LASSERT(ea2 != NULL);
1777 memcpy(ea2, ea, body->eadatasize);
1779 if (body->valid & OBD_MD_FLCOOKIE) {
1780 LASSERT(master->rq_repmsg->buflens[2] >=
1781 slave->rq_repmsg->buflens[2]);
1782 cookie = lustre_msg_buf(slave->rq_repmsg, 2,
1783 slave->rq_repmsg->buflens[2]);
1784 LASSERT(cookie != NULL);
1786 cookie2 = lustre_msg_buf(master->rq_repmsg, 2,
1787 master->rq_repmsg->buflens[2]);
1788 LASSERT(cookie2 != NULL);
1789 memcpy(cookie2, cookie, slave->rq_repmsg->buflens[2]);
1790 body2->valid |= OBD_MD_FLCOOKIE;
1795 static int mds_reint_unlink_remote(struct mds_update_record *rec,
1796 int offset, struct ptlrpc_request *req,
1797 struct lustre_handle *parent_lockh,
1798 int update_mode, struct dentry *dparent,
1799 struct lustre_handle *child_lockh,
1800 struct dentry *dchild)
1802 struct obd_device *obd = req->rq_export->exp_obd;
1803 struct mds_obd *mds = mds_req2mds(req);
1804 struct ptlrpc_request *request = NULL;
1805 int rc = 0, cleanup_phase = 0;
1806 struct mdc_op_data op_data;
1810 LASSERT(offset == 1 || offset == 3);
1812 /* time to drop i_nlink on remote MDS */
1813 memset(&op_data, 0, sizeof(op_data));
1814 mds_pack_dentry2id(obd, &op_data.id1, dchild, 1);
1815 op_data.create_mode = rec->ur_mode;
1817 DEBUG_REQ(D_INODE, req, "unlink %*s (remote inode "DLID4")",
1818 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data.id1));
1820 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1821 DEBUG_REQ(D_HA, req, "unlink %*s (remote inode "DLID4")",
1822 rec->ur_namelen - 1, rec->ur_name, OLID4(&op_data.id1));
1825 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1826 op_data.create_mode |= MDS_MODE_REPLAY;
1828 rc = md_unlink(mds->mds_lmv_exp, &op_data, &request);
1833 mds_copy_unlink_reply(req, request);
1834 ptlrpc_req_finished(request);
1838 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
1841 GOTO(cleanup, rc = PTR_ERR(handle));
1842 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, dchild);
1843 rc = mds_finish_transno(mds, dparent->d_inode, handle, req,
1848 req->rq_status = rc;
1851 if (parent_lockh[1].cookie != 0)
1852 ldlm_lock_decref(parent_lockh + 1, update_mode);
1854 ldlm_lock_decref(child_lockh, LCK_EX);
1856 ldlm_lock_decref(parent_lockh, LCK_PW);
1858 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
1865 static int mds_reint_unlink(struct mds_update_record *rec, int offset,
1866 struct ptlrpc_request *req,
1867 struct lustre_handle *lh)
1869 struct dentry *dparent = NULL, *dchild;
1870 struct mds_obd *mds = mds_req2mds(req);
1871 struct obd_device *obd = req->rq_export->exp_obd;
1872 struct mds_body *body = NULL;
1873 struct inode *child_inode = NULL;
1874 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1875 struct lustre_handle child_lockh = {0};
1876 struct lustre_handle child_reuse_lockh = {0};
1877 struct lustre_handle *slave_lockh = NULL;
1878 char idname[LL_ID_NAMELEN];
1879 struct llog_create_locks *lcl = NULL;
1880 void *handle = NULL;
1881 int rc = 0, cleanup_phase = 0;
1882 int unlink_by_id = 0;
1886 LASSERT(offset == 1 || offset == 3);
1888 DEBUG_REQ(D_INODE, req, "parent ino "LPU64"/%u, child %s",
1889 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
1892 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1894 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1895 DEBUG_REQ(D_HA, req, "unlink replay\n");
1896 LASSERT(offset == 1); /* should not come from intent */
1897 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1898 lustre_msg_buf(req->rq_reqmsg, offset + 2, 0),
1899 req->rq_repmsg->buflens[2]);
1902 MD_COUNTER_INCREMENT(obd, unlink);
1904 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1905 GOTO(cleanup, rc = -ENOENT);
1907 if (rec->ur_namelen == 1) {
1908 /* this is request to drop i_nlink on local inode */
1910 rec->ur_name = idname;
1911 rc = mds_create_local_dentry(rec, obd);
1912 if (rc == -ENOENT || (rec->ur_mode & MDS_MODE_REPLAY)) {
1913 DEBUG_REQ(D_HA, req,
1914 "drop nlink on inode "DLID4" (replay)",
1915 OLID4(rec->ur_id1));
1921 if (rec->ur_mode & MDS_MODE_DONT_LOCK) {
1922 /* master mds for directory asks slave removing inode is already
1924 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL,
1925 LCK_PW, parent_lockh,
1926 &update_mode, rec->ur_name,
1928 MDS_INODELOCK_UPDATE);
1929 if (IS_ERR(dparent))
1930 GOTO(cleanup, rc = PTR_ERR(dparent));
1931 dchild = ll_lookup_one_len(rec->ur_name, dparent,
1932 rec->ur_namelen - 1);
1934 GOTO(cleanup, rc = PTR_ERR(dchild));
1935 child_lockh.cookie = 0;
1936 LASSERT(!(dchild->d_flags & DCACHE_CROSS_REF));
1937 LASSERT(dchild->d_inode != NULL);
1938 LASSERT(S_ISDIR(dchild->d_inode->i_mode));
1940 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1941 parent_lockh, &dparent,
1942 LCK_PW, MDS_INODELOCK_UPDATE,
1943 &update_mode, rec->ur_name,
1944 rec->ur_namelen, &child_lockh,
1946 MDS_INODELOCK_LOOKUP |
1947 MDS_INODELOCK_UPDATE);
1952 if (dchild->d_flags & DCACHE_CROSS_REF) {
1953 /* we should have parent lock only here */
1954 LASSERT(unlink_by_id == 0);
1955 LASSERT(dchild->d_mdsnum != mds->mds_num);
1956 mds_reint_unlink_remote(rec, offset, req, parent_lockh,
1957 update_mode, dparent, &child_lockh, dchild);
1961 cleanup_phase = 1; /* dchild, dparent, locks */
1964 child_inode = dchild->d_inode;
1965 if (child_inode == NULL) {
1966 CDEBUG(D_INODE, "child doesn't exist (dir %lu, name %s)\n",
1967 dparent ? dparent->d_inode->i_ino : 0, rec->ur_name);
1968 GOTO(cleanup, rc = -ENOENT);
1971 cleanup_phase = 2; /* dchild has a lock */
1973 /* We have to do these checks ourselves, in case we are making an
1974 * orphan. The client tells us whether rmdir() or unlink() was called,
1975 * so we need to return appropriate errors (bug 72).
1977 * We don't have to check permissions, because vfs_rename (called from
1978 * mds_open_unlink_rename) also calls may_delete. */
1979 if ((rec->ur_mode & S_IFMT) == S_IFDIR) {
1980 if (!S_ISDIR(child_inode->i_mode))
1981 GOTO(cleanup, rc = -ENOTDIR);
1983 if (S_ISDIR(child_inode->i_mode))
1984 GOTO(cleanup, rc = -EISDIR);
1987 /* handle splitted dir */
1988 rc = mds_lock_slave_objs(obd, dchild, &slave_lockh);
1992 /* Step 4: Get a lock on the ino to sync with creation WRT inode
1993 * reuse (see bug 2029). */
1994 rc = mds_lock_new_child(obd, child_inode, &child_reuse_lockh);
1997 cleanup_phase = 3; /* child inum lock */
1999 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dparent->d_inode->i_sb);
2001 /* ldlm_reply in buf[0] if called via intent */
2007 body = lustre_msg_buf(req->rq_repmsg, offset, sizeof (*body));
2008 LASSERT(body != NULL);
2010 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
2011 DOWN_READ_I_ALLOC_SEM(child_inode);
2012 cleanup_phase = 4; /* up(&child_inode->i_sem) when finished */
2014 /* If this is potentially the last reference to this inode, get the
2015 * OBD EA data first so the client can destroy OST objects. We
2016 * only do the object removal later if no open files/links remain. */
2017 if ((S_ISDIR(child_inode->i_mode) && child_inode->i_nlink == 2) ||
2018 child_inode->i_nlink == 1) {
2019 if (mds_orphan_open_count(child_inode) > 0) {
2020 /* need to lock pending_dir before transaction */
2021 down(&mds->mds_pending_dir->d_inode->i_sem);
2022 cleanup_phase = 5; /* up(&pending_dir->i_sem) */
2023 } else if (S_ISREG(child_inode->i_mode)) {
2024 mds_pack_inode2body(obd, body, child_inode, 0);
2025 mds_pack_md(obd, req->rq_repmsg, offset + 1,
2026 body, child_inode, MDS_PACK_MD_LOCK);
2030 /* Step 4: Do the unlink: we already verified ur_mode above (bug 72) */
2031 switch (child_inode->i_mode & S_IFMT) {
2033 /* Drop any lingering child directories before we start our
2034 * transaction, to avoid doing multiple inode dirty/delete
2035 * in our compound transaction (bug 1321). */
2036 shrink_dcache_parent(dchild);
2037 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_RMDIR,
2040 GOTO(cleanup, rc = PTR_ERR(handle));
2041 rc = vfs_rmdir(dparent->d_inode, dchild);
2044 #warning "optimization is possible here: we could drop nlink w/o removing local dentry in FIDS/"
2045 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_repmsg,
2047 handle = fsfilt_start_log(obd, dparent->d_inode,
2048 FSFILT_OP_UNLINK, NULL,
2049 le32_to_cpu(lmm->lmm_stripe_count));
2051 GOTO(cleanup, rc = PTR_ERR(handle));
2052 rc = vfs_unlink(dparent->d_inode, dchild);
2060 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK,
2063 GOTO(cleanup, rc = PTR_ERR(handle));
2064 rc = vfs_unlink(dparent->d_inode, dchild);
2067 CERROR("bad file type %o unlinking %s\n", rec->ur_mode,
2070 GOTO(cleanup, rc = -EINVAL);
2073 if (rc == 0 && child_inode->i_nlink == 0) {
2074 if (mds_orphan_open_count(child_inode) > 0)
2075 rc = mds_orphan_add_link(rec, obd, dchild);
2078 GOTO(cleanup, rc = 0);
2080 if (!S_ISREG(child_inode->i_mode))
2083 if (!(body->valid & OBD_MD_FLEASIZE)) {
2084 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
2085 OBD_MD_FLATIME | OBD_MD_FLMTIME);
2086 } else if (mds_log_op_unlink(obd, child_inode,
2087 lustre_msg_buf(req->rq_repmsg, offset + 1, 0),
2088 req->rq_repmsg->buflens[offset + 1],
2089 lustre_msg_buf(req->rq_repmsg, offset + 2, 0),
2090 req->rq_repmsg->buflens[offset+2],
2092 body->valid |= OBD_MD_FLCOOKIE;
2103 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
2104 LTIME_S(iattr.ia_mtime) = rec->ur_time;
2105 LTIME_S(iattr.ia_ctime) = rec->ur_time;
2107 err = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
2109 CERROR("error on parent setattr: rc = %d\n", err);
2111 rc = mds_finish_transno(mds, dparent ? dparent->d_inode : NULL,
2112 handle, req, rc, 0);
2114 (void)obd_set_info(mds->mds_lov_exp, strlen("unlinked"),
2115 "unlinked", 0, NULL);
2116 switch(cleanup_phase) {
2117 case 5: /* pending_dir semaphore */
2118 up(&mds->mds_pending_dir->d_inode->i_sem);
2119 case 4: /* child inode semaphore */
2120 UP_READ_I_ALLOC_SEM(child_inode);
2121 /* handle splitted dir */
2123 /* master directory can be non-empty or something else ... */
2124 mds_unlink_slave_objs(obd, dchild);
2127 ptlrpc_save_llog_lock(req, lcl);
2128 case 3: /* child ino-reuse lock */
2129 if (rc && body != NULL) {
2130 // Don't unlink the OST objects if the MDS unlink failed
2134 ldlm_lock_decref(&child_reuse_lockh, LCK_EX);
2136 ptlrpc_save_lock(req, &child_reuse_lockh, LCK_EX);
2137 case 2: /* child lock */
2138 mds_unlock_slave_objs(obd, dchild, slave_lockh);
2139 if (child_lockh.cookie)
2140 ldlm_lock_decref(&child_lockh, LCK_EX);
2141 case 1: /* child and parent dentry, parent lock */
2143 if (parent_lockh[1].cookie != 0)
2144 ldlm_lock_decref(parent_lockh + 1, update_mode);
2147 ldlm_lock_decref(parent_lockh, LCK_PW);
2149 ptlrpc_save_lock(req, parent_lockh, LCK_PW);
2156 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2159 req->rq_status = rc;
2164 * to service requests from remote MDS to increment i_nlink
2166 static int mds_reint_link_acquire(struct mds_update_record *rec,
2167 int offset, struct ptlrpc_request *req,
2168 struct lustre_handle *lh)
2170 struct obd_device *obd = req->rq_export->exp_obd;
2171 struct ldlm_res_id src_res_id = { .name = {0} };
2172 struct lustre_handle *handle = NULL, src_lockh = {0};
2173 struct mds_obd *mds = mds_req2mds(req);
2174 int rc = 0, cleanup_phase = 0;
2175 struct dentry *de_src = NULL;
2176 ldlm_policy_data_t policy;
2180 DEBUG_REQ(D_INODE, req, "%s: request to acquire i_nlinks "DLID4"\n",
2181 obd->obd_name, OLID4(rec->ur_id1));
2183 /* Step 1: Lookup the source inode and target directory by ID */
2184 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2186 GOTO(cleanup, rc = PTR_ERR(de_src));
2187 cleanup_phase = 1; /* source dentry */
2189 src_res_id.name[0] = id_fid(rec->ur_id1);
2190 src_res_id.name[1] = id_group(rec->ur_id1);
2191 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
2193 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2194 src_res_id, LDLM_IBITS, &policy,
2195 LCK_EX, &flags, mds_blocking_ast,
2196 ldlm_completion_ast, NULL, NULL,
2197 NULL, 0, NULL, &src_lockh);
2199 GOTO(cleanup, rc = -ENOLCK);
2200 cleanup_phase = 2; /* lock */
2202 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2204 handle = fsfilt_start(obd, de_src->d_inode, FSFILT_OP_LINK, NULL);
2205 if (IS_ERR(handle)) {
2206 rc = PTR_ERR(handle);
2209 de_src->d_inode->i_nlink++;
2210 mark_inode_dirty(de_src->d_inode);
2213 rc = mds_finish_transno(mds, de_src ? de_src->d_inode : NULL,
2214 handle, req, rc, 0);
2216 switch (cleanup_phase) {
2219 ldlm_lock_decref(&src_lockh, LCK_EX);
2221 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2227 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2230 req->rq_status = rc;
2235 * request to link to foreign inode:
2236 * - acquire i_nlinks on this inode
2239 static int mds_reint_link_to_remote(struct mds_update_record *rec,
2240 int offset, struct ptlrpc_request *req,
2241 struct lustre_handle *lh)
2243 struct lustre_handle *handle = NULL, tgt_dir_lockh[2] = {{0}, {0}};
2244 struct obd_device *obd = req->rq_export->exp_obd;
2245 struct dentry *de_tgt_dir = NULL;
2246 struct mds_obd *mds = mds_req2mds(req);
2247 int rc = 0, cleanup_phase = 0;
2248 struct mdc_op_data op_data;
2249 struct ptlrpc_request *request = NULL;
2253 DEBUG_REQ(D_INODE, req, "%s: request to link "DLID4
2254 ":%*s to foreign inode "DLID4"\n", obd->obd_name,
2255 OLID4(rec->ur_id2), rec->ur_namelen - 1, rec->ur_name,
2256 OLID4(rec->ur_id1));
2258 de_tgt_dir = mds_id2locked_dentry(obd, rec->ur_id2, NULL, LCK_EX,
2259 tgt_dir_lockh, &update_mode,
2260 rec->ur_name, rec->ur_namelen - 1,
2261 MDS_INODELOCK_UPDATE);
2262 if (IS_ERR(de_tgt_dir))
2263 GOTO(cleanup, rc = PTR_ERR(de_tgt_dir));
2266 op_data.id1 = *(rec->ur_id1);
2267 op_data.namelen = 0;
2268 op_data.name = NULL;
2269 rc = md_link(mds->mds_lmv_exp, &op_data, &request);
2275 ptlrpc_req_finished(request);
2277 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_tgt_dir->d_inode->i_sb);
2279 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2280 if (IS_ERR(handle)) {
2281 rc = PTR_ERR(handle);
2285 rc = fsfilt_add_dir_entry(obd, de_tgt_dir, rec->ur_name,
2286 rec->ur_namelen - 1, id_ino(rec->ur_id1),
2287 id_gen(rec->ur_id1), id_group(rec->ur_id1),
2288 id_fid(rec->ur_id1));
2292 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2293 handle, req, rc, 0);
2296 switch (cleanup_phase) {
2299 /* FIXME: drop i_nlink on remote inode here */
2300 CERROR("MUST drop drop i_nlink here\n");
2305 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2307 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2310 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2312 ptlrpc_save_lock(req, tgt_dir_lockh+1, update_mode);
2318 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2321 req->rq_status = rc;
2325 static int mds_reint_link(struct mds_update_record *rec, int offset,
2326 struct ptlrpc_request *req,
2327 struct lustre_handle *lh)
2329 struct obd_device *obd = req->rq_export->exp_obd;
2330 struct dentry *de_src = NULL;
2331 struct dentry *de_tgt_dir = NULL;
2332 struct dentry *dchild = NULL;
2333 struct mds_obd *mds = mds_req2mds(req);
2334 struct lustre_handle *handle = NULL;
2335 struct lustre_handle tgt_dir_lockh[2] = {{0}, {0}}, src_lockh = {0};
2336 struct ldlm_res_id src_res_id = { .name = {0} };
2337 struct ldlm_res_id tgt_dir_res_id = { .name = {0} };
2338 ldlm_policy_data_t src_policy ={.l_inodebits = {MDS_INODELOCK_UPDATE}};
2339 ldlm_policy_data_t tgt_dir_policy =
2340 {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2341 int rc = 0, cleanup_phase = 0;
2343 int update_mode = 0;
2347 LASSERT(offset == 1);
2349 DEBUG_REQ(D_INODE, req, "original "LPU64"/%u to "LPU64"/%u %s",
2350 id_ino(rec->ur_id1), id_gen(rec->ur_id1),
2351 id_ino(rec->ur_id2), id_gen(rec->ur_id2),
2354 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
2355 MD_COUNTER_INCREMENT(obd, link);
2357 // memset(tgt_dir_lockh, 0, 2*sizeof(tgt_dir_lockh[0]));
2358 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
2359 GOTO(cleanup, rc = -ENOENT);
2361 if (id_group(rec->ur_id1) != mds->mds_num) {
2362 rc = mds_reint_link_to_remote(rec, offset, req, lh);
2366 if (rec->ur_namelen == 1) {
2367 rc = mds_reint_link_acquire(rec, offset, req, lh);
2371 /* Step 1: Lookup the source inode and target directory by ID */
2372 de_src = mds_id2dentry(obd, rec->ur_id1, NULL);
2374 GOTO(cleanup, rc = PTR_ERR(de_src));
2376 cleanup_phase = 1; /* source dentry */
2378 de_tgt_dir = mds_id2dentry(obd, rec->ur_id2, NULL);
2379 if (IS_ERR(de_tgt_dir)) {
2380 rc = PTR_ERR(de_tgt_dir);
2385 cleanup_phase = 2; /* target directory dentry */
2387 CDEBUG(D_INODE, "linking %*s/%s to inode %lu\n",
2388 de_tgt_dir->d_name.len, de_tgt_dir->d_name.name,
2389 rec->ur_name, de_src->d_inode->i_ino);
2391 /* Step 2: Take the two locks */
2392 src_res_id.name[0] = id_fid(rec->ur_id1);
2393 src_res_id.name[1] = id_group(rec->ur_id1);
2394 tgt_dir_res_id.name[0] = id_fid(rec->ur_id2);
2395 tgt_dir_res_id.name[1] = id_group(rec->ur_id2);
2398 if (IS_PDIROPS(de_tgt_dir->d_inode)) {
2400 update_mode = mds_lock_mode_for_dir(obd, de_tgt_dir, LCK_EX);
2402 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
2403 tgt_dir_res_id, LDLM_IBITS,
2404 &src_policy, update_mode, &flags,
2406 ldlm_completion_ast, NULL, NULL,
2407 NULL, 0, NULL, tgt_dir_lockh + 1);
2409 GOTO(cleanup, rc = -ENOLCK);
2412 tgt_dir_res_id.name[2] = full_name_hash(rec->ur_name,
2413 rec->ur_namelen - 1);
2414 CDEBUG(D_INFO, "take lock on %lu:%lu:"LPX64"\n",
2415 (unsigned long)id_fid(rec->ur_id2),
2416 (unsigned long)id_group(rec->ur_id2),
2417 tgt_dir_res_id.name[2]);
2420 rc = enqueue_ordered_locks(obd, &src_res_id, &src_lockh, LCK_EX,
2421 &src_policy, &tgt_dir_res_id, tgt_dir_lockh,
2422 LCK_EX, &tgt_dir_policy);
2426 cleanup_phase = 3; /* locks */
2428 /* Step 3: Lookup the child */
2429 dchild = ll_lookup_one_len(rec->ur_name, de_tgt_dir,
2430 rec->ur_namelen - 1);
2431 if (IS_ERR(dchild)) {
2432 rc = PTR_ERR(dchild);
2433 if (rc != -EPERM && rc != -EACCES)
2434 CERROR("child lookup error %d\n", rc);
2438 cleanup_phase = 4; /* child dentry */
2440 if (dchild->d_inode) {
2441 CDEBUG(D_INODE, "child exists (dir %lu, name %s)\n",
2442 de_tgt_dir->d_inode->i_ino, rec->ur_name);
2447 /* Step 4: Do it. */
2448 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE, de_src->d_inode->i_sb);
2450 handle = fsfilt_start(obd, de_tgt_dir->d_inode, FSFILT_OP_LINK, NULL);
2451 if (IS_ERR(handle)) {
2452 rc = PTR_ERR(handle);
2456 rc = vfs_link(de_src, de_tgt_dir->d_inode, dchild);
2457 if (rc && rc != -EPERM && rc != -EACCES)
2458 CERROR("vfs_link error %d\n", rc);
2460 rc = mds_finish_transno(mds, de_tgt_dir ? de_tgt_dir->d_inode : NULL,
2461 handle, req, rc, 0);
2464 switch (cleanup_phase) {
2465 case 4: /* child dentry */
2469 ldlm_lock_decref(&src_lockh, LCK_EX);
2470 ldlm_lock_decref(tgt_dir_lockh, LCK_EX);
2472 ptlrpc_save_lock(req, &src_lockh, LCK_EX);
2473 ptlrpc_save_lock(req, tgt_dir_lockh, LCK_EX);
2475 case 2: /* target dentry */
2477 if (tgt_dir_lockh[1].cookie && update_mode)
2478 ldlm_lock_decref(tgt_dir_lockh + 1, update_mode);
2482 case 1: /* source dentry */
2487 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2490 req->rq_status = rc;
2494 /* The idea here is that we need to get four locks in the end:
2495 * one on each parent directory, one on each child. We need to take
2496 * these locks in some kind of order (to avoid deadlocks), and the order
2497 * I selected is "increasing resource number" order. We need to look up
2498 * the children, however, before we know what the resource number(s) are.
2499 * Thus the following plan:
2501 * 1,2. Look up the parents
2502 * 3,4. Look up the children
2503 * 5. Take locks on the parents and children, in order
2504 * 6. Verify that the children haven't changed since they were looked up
2506 * If there was a race and the children changed since they were first looked
2507 * up, it is possible that mds_verify_child() will be able to just grab the
2508 * lock on the new child resource (if it has a higher resource than any other)
2509 * but we need to compare against not only its parent, but also against the
2510 * parent and child of the "other half" of the rename, hence maxres_{src,tgt}.
2512 * We need the fancy igrab() on the child inodes because we aren't holding a
2513 * lock on the parent after the lookup is done, so dentry->d_inode may change
2514 * at any time, and igrab() itself doesn't like getting passed a NULL argument.
2516 static int mds_get_parents_children_locked(struct obd_device *obd,
2517 struct mds_obd *mds,
2518 struct lustre_id *p1_id,
2519 struct dentry **de_srcdirp,
2520 struct lustre_id *p2_id,
2521 struct dentry **de_tgtdirp,
2523 const char *old_name, int old_len,
2524 struct dentry **de_oldp,
2525 const char *new_name, int new_len,
2526 struct dentry **de_newp,
2527 struct lustre_handle *dlm_handles,
2530 struct ldlm_res_id p1_res_id = { .name = {0} };
2531 struct ldlm_res_id p2_res_id = { .name = {0} };
2532 struct ldlm_res_id c1_res_id = { .name = {0} };
2533 struct ldlm_res_id c2_res_id = { .name = {0} };
2534 ldlm_policy_data_t p_policy = {.l_inodebits = {MDS_INODELOCK_UPDATE}};
2535 /* Only dentry should change, but the inode itself would be
2537 ldlm_policy_data_t c1_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP}};
2538 /* If something is going to be replaced, both dentry and inode locks are
2540 ldlm_policy_data_t c2_policy = {.l_inodebits = {MDS_INODELOCK_LOOKUP|
2541 MDS_INODELOCK_UPDATE}};
2542 struct ldlm_res_id *maxres_src, *maxres_tgt;
2543 struct inode *inode;
2544 __u32 child1_gen = 0;
2545 __u32 child2_gen = 0;
2546 unsigned long child1_ino;
2547 unsigned long child2_ino;
2548 int rc = 0, cleanup_phase = 0;
2551 /* Step 1: Lookup the source directory */
2552 *de_srcdirp = mds_id2dentry(obd, p1_id, NULL);
2553 if (IS_ERR(*de_srcdirp))
2554 GOTO(cleanup, rc = PTR_ERR(*de_srcdirp));
2556 cleanup_phase = 1; /* source directory dentry */
2558 p1_res_id.name[0] = id_fid(p1_id);
2559 p1_res_id.name[1] = id_group(p1_id);
2561 /* Step 2: Lookup the target directory */
2562 if (id_equal_stc(p1_id, p2_id)) {
2563 *de_tgtdirp = dget(*de_srcdirp);
2565 *de_tgtdirp = mds_id2dentry(obd, p2_id, NULL);
2566 if (IS_ERR(*de_tgtdirp)) {
2567 rc = PTR_ERR(*de_tgtdirp);
2573 cleanup_phase = 2; /* target directory dentry */
2575 p2_res_id.name[0] = id_fid(p2_id);
2576 p2_res_id.name[1] = id_group(p2_id);
2579 dlm_handles[5].cookie = 0;
2580 dlm_handles[6].cookie = 0;
2582 if (IS_PDIROPS((*de_srcdirp)->d_inode)) {
2584 * get a temp lock on just fid, group to flush client cache and
2585 * to protect dirs from concurrent splitting.
2587 rc = enqueue_ordered_locks(obd, &p1_res_id, &dlm_handles[5],
2588 LCK_PW, &p_policy, &p2_res_id,
2589 &dlm_handles[6], LCK_PW, &p_policy);
2593 p1_res_id.name[2] = full_name_hash(old_name, old_len - 1);
2594 p2_res_id.name[2] = full_name_hash(new_name, new_len - 1);
2596 CDEBUG(D_INFO, "take locks on "
2597 LPX64":"LPX64":"LPX64", "LPX64":"LPX64":"LPX64"\n",
2598 p1_res_id.name[0], p1_res_id.name[1], p1_res_id.name[2],
2599 p2_res_id.name[0], p2_res_id.name[1], p2_res_id.name[2]);
2604 /* Step 3: Lookup the source child entry */
2605 *de_oldp = ll_lookup_one_len(old_name, *de_srcdirp,
2607 if (IS_ERR(*de_oldp)) {
2608 rc = PTR_ERR(*de_oldp);
2609 CERROR("old child lookup error (%*s): %d\n",
2610 old_len - 1, old_name, rc);
2614 cleanup_phase = 4; /* original name dentry */
2616 inode = (*de_oldp)->d_inode;
2617 if (inode != NULL) {
2618 struct lustre_id sid;
2620 inode = igrab(inode);
2622 GOTO(cleanup, rc = -ENOENT);
2624 down(&inode->i_sem);
2625 rc = mds_read_inode_sid(obd, inode, &sid);
2628 CERROR("Can't read inode self id, inode %lu, "
2629 "rc %d\n", inode->i_ino, rc);
2634 c1_res_id.name[0] = id_fid(&sid);
2635 c1_res_id.name[1] = id_group(&sid);
2636 child1_gen = inode->i_generation;
2637 child1_ino = inode->i_ino;
2639 } else if ((*de_oldp)->d_flags & DCACHE_CROSS_REF) {
2640 c1_res_id.name[0] = (*de_oldp)->d_fid;
2641 c1_res_id.name[1] = (*de_oldp)->d_mdsnum;
2642 child1_gen = (*de_oldp)->d_generation;
2643 child1_ino = (*de_oldp)->d_inum;
2645 GOTO(cleanup, rc = -ENOENT);
2648 /* Step 4: Lookup the target child entry */
2649 *de_newp = ll_lookup_one_len(new_name, *de_tgtdirp,
2651 if (IS_ERR(*de_newp)) {
2652 rc = PTR_ERR(*de_newp);
2653 CERROR("new child lookup error (%*s): %d\n",
2654 old_len - 1, old_name, rc);
2658 cleanup_phase = 5; /* target dentry */
2660 inode = (*de_newp)->d_inode;
2661 if (inode != NULL) {
2662 struct lustre_id sid;
2664 inode = igrab(inode);
2668 down(&inode->i_sem);
2669 rc = mds_read_inode_sid(obd, inode, &sid);
2672 CERROR("Can't read inode self id, inode %lu, "
2673 "rc %d\n", inode->i_ino, rc);
2677 c2_res_id.name[0] = id_fid(&sid);
2678 c2_res_id.name[1] = id_group(&sid);
2679 child2_gen = inode->i_generation;
2680 child2_ino = inode->i_ino;
2682 } else if ((*de_newp)->d_flags & DCACHE_CROSS_REF) {
2683 c2_res_id.name[0] = (*de_newp)->d_fid;
2684 c2_res_id.name[1] = (*de_newp)->d_mdsnum;
2685 child2_gen = (*de_newp)->d_generation;
2686 child2_ino = (*de_newp)->d_inum;
2690 /* Step 5: Take locks on the parents and child(ren) */
2691 maxres_src = &p1_res_id;
2692 maxres_tgt = &p2_res_id;
2693 cleanup_phase = 5; /* target dentry */
2695 if (c1_res_id.name[0] != 0 && res_gt(&c1_res_id, &p1_res_id, NULL, NULL))
2696 maxres_src = &c1_res_id;
2697 if (c2_res_id.name[0] != 0 && res_gt(&c2_res_id, &p2_res_id, NULL, NULL))
2698 maxres_tgt = &c2_res_id;
2700 rc = enqueue_4ordered_locks(obd, &p1_res_id, &dlm_handles[0], parent_mode,
2702 &p2_res_id, &dlm_handles[1], parent_mode,
2704 &c1_res_id, &dlm_handles[2], child_mode,
2706 &c2_res_id, &dlm_handles[3], child_mode,
2711 cleanup_phase = 6; /* parent and child(ren) locks */
2713 /* Step 6a: Re-lookup source child to verify it hasn't changed */
2714 rc = mds_verify_child(obd, &p1_res_id, &dlm_handles[0], *de_srcdirp,
2715 parent_mode, &c1_res_id, &dlm_handles[2],
2716 de_oldp, child_mode, &c1_policy, old_name,old_len,
2717 maxres_tgt, child1_ino, child1_gen);
2719 if (c2_res_id.name[0] != 0)
2720 ldlm_lock_decref(&dlm_handles[3], child_mode);
2721 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2728 if (!DENTRY_VALID(*de_oldp))
2729 GOTO(cleanup, rc = -ENOENT);
2731 /* Step 6b: Re-lookup target child to verify it hasn't changed */
2732 rc = mds_verify_child(obd, &p2_res_id, &dlm_handles[1], *de_tgtdirp,
2733 parent_mode, &c2_res_id, &dlm_handles[3],
2734 de_newp, child_mode, &c2_policy, new_name,
2735 new_len, maxres_src, child2_ino, child2_gen);
2737 ldlm_lock_decref(&dlm_handles[2], child_mode);
2738 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2748 switch (cleanup_phase) {
2749 case 6: /* child lock(s) */
2750 if (c2_res_id.name[0] != 0)
2751 ldlm_lock_decref(&dlm_handles[3], child_mode);
2752 if (c1_res_id.name[0] != 0)
2753 ldlm_lock_decref(&dlm_handles[2], child_mode);
2754 if (dlm_handles[1].cookie != 0)
2755 ldlm_lock_decref(&dlm_handles[1], parent_mode);
2756 if (dlm_handles[0].cookie != 0)
2757 ldlm_lock_decref(&dlm_handles[0], parent_mode);
2758 case 5: /* target dentry */
2760 case 4: /* source dentry */
2764 if (dlm_handles[5].cookie != 0)
2765 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
2766 if (dlm_handles[6].cookie != 0)
2767 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
2769 case 2: /* target directory dentry */
2770 l_dput(*de_tgtdirp);
2771 case 1: /* source directry dentry */
2772 l_dput(*de_srcdirp);
2780 * checks if dentry can be removed. This function also handles cross-ref
2783 static int mds_check_for_rename(struct obd_device *obd,
2784 struct dentry *dentry)
2786 struct mds_obd *mds = &obd->u.mds;
2787 struct lustre_handle *rlockh;
2788 struct ptlrpc_request *req;
2789 struct mdc_op_data op_data;
2790 struct lookup_intent it;
2791 int handle_size, rc = 0;
2794 LASSERT(dentry != NULL);
2796 if (dentry->d_inode) {
2797 if (S_ISDIR(dentry->d_inode->i_mode) &&
2798 !mds_is_dir_empty(obd, dentry))
2801 LASSERT((dentry->d_flags & DCACHE_CROSS_REF));
2802 handle_size = sizeof(struct lustre_handle);
2804 OBD_ALLOC(rlockh, handle_size);
2808 memset(rlockh, 0, handle_size);
2809 memset(&op_data, 0, sizeof(op_data));
2810 mds_pack_dentry2id(obd, &op_data.id1, dentry, 1);
2812 it.it_op = IT_UNLINK;
2813 rc = md_enqueue(mds->mds_lmv_exp, LDLM_IBITS, &it, LCK_EX,
2814 &op_data, rlockh, NULL, 0, ldlm_completion_ast,
2815 mds_blocking_ast, NULL);
2820 if (rlockh->cookie != 0)
2821 ldlm_lock_decref(rlockh, LCK_EX);
2823 if (it.d.lustre.it_data) {
2824 req = (struct ptlrpc_request *)it.d.lustre.it_data;
2825 ptlrpc_req_finished(req);
2828 if (it.d.lustre.it_status)
2829 rc = it.d.lustre.it_status;
2830 OBD_FREE(rlockh, handle_size);
2835 static int mds_add_local_dentry(struct mds_update_record *rec, int offset,
2836 struct ptlrpc_request *req, struct lustre_id *id,
2837 struct dentry *de_dir, struct dentry *de,
2840 struct obd_device *obd = req->rq_export->exp_obd;
2841 struct mds_obd *mds = mds_req2mds(req);
2842 void *handle = NULL;
2848 * name exists and points to local inode try to unlink this name
2849 * and create new one.
2851 CDEBUG(D_OTHER, "%s: %s points to local inode %lu/%lu\n",
2852 obd->obd_name, rec->ur_tgt, (unsigned long)de->d_inode->i_ino,
2853 (unsigned long)de->d_inode->i_generation);
2855 /* checking if we can remove local dentry. */
2856 rc = mds_check_for_rename(obd, de);
2860 handle = fsfilt_start(obd, de_dir->d_inode,
2861 FSFILT_OP_RENAME, NULL);
2863 GOTO(cleanup, rc = PTR_ERR(handle));
2864 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
2867 } else if (de->d_flags & DCACHE_CROSS_REF) {
2868 struct lustre_id de_id;
2870 /* name exists and points to remote inode */
2871 mds_pack_dentry2id(obd, &de_id, de, 1);
2873 CDEBUG(D_OTHER, "%s: %s points to remote inode "DLID4"\n",
2874 obd->obd_name, rec->ur_tgt, OLID4(&de_id));
2876 /* checking if we can remove local dentry. */
2877 rc = mds_check_for_rename(obd, de);
2882 * to be fully POSIX compatible, we should add one more check:
2884 * if de_new is subdir of dir rec->ur_id1. If so - return
2887 * I do not know how to implement it right now, because
2888 * inodes/dentries for new and old names lie on different MDS,
2889 * so add this notice here just to make it visible for the rest
2890 * of developers and do not forget about. And when this check
2891 * will be added, del_cross_ref should gone, that is local
2892 * dentry is able to be removed if all checks passed.
2894 * Currently -EEXISTS is returned by fsfilt_add_dir_entry() what
2895 * is not fully correct. --umka
2898 if (del_cross_ref) {
2899 handle = fsfilt_start(obd, de_dir->d_inode,
2900 FSFILT_OP_RENAME, NULL);
2902 GOTO(cleanup, rc = PTR_ERR(handle));
2903 rc = fsfilt_del_dir_entry(req->rq_export->exp_obd, de);
2909 /* name doesn't exist. the simplest case. */
2910 handle = fsfilt_start(obd, de_dir->d_inode,
2911 FSFILT_OP_LINK, NULL);
2913 GOTO(cleanup, rc = PTR_ERR(handle));
2916 rc = fsfilt_add_dir_entry(obd, de_dir, rec->ur_tgt,
2917 rec->ur_tgtlen - 1, id_ino(id),
2918 id_gen(id), id_group(id), id_fid(id));
2920 CERROR("add_dir_entry() returned error %d\n", rc);
2926 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
2927 handle, req, rc, 0);
2932 static int mds_del_local_dentry(struct mds_update_record *rec, int offset,
2933 struct ptlrpc_request *req, struct dentry *de_dir,
2936 struct obd_device *obd = req->rq_export->exp_obd;
2937 struct mds_obd *mds = mds_req2mds(req);
2938 void *handle = NULL;
2942 handle = fsfilt_start(obd, de_dir->d_inode, FSFILT_OP_UNLINK, NULL);
2944 GOTO(cleanup, rc = PTR_ERR(handle));
2945 rc = fsfilt_del_dir_entry(obd, de);
2950 rc = mds_finish_transno(mds, de_dir ? de_dir->d_inode : NULL,
2951 handle, req, rc, 0);
2955 static int mds_reint_rename_create_name(struct mds_update_record *rec,
2956 int offset, struct ptlrpc_request *req)
2958 struct lustre_handle parent_lockh[2] = {{0}, {0}};
2959 struct obd_device *obd = req->rq_export->exp_obd;
2960 struct mds_obd *mds = mds_req2mds(req);
2961 struct lustre_handle child_lockh = {0};
2962 struct dentry *de_tgtdir = NULL;
2963 struct dentry *de_new = NULL;
2964 int cleanup_phase = 0;
2965 int update_mode, rc = 0;
2969 * another MDS executing rename operation has asked us to create target
2970 * name. such a creation should destroy existing target name.
2972 CDEBUG(D_OTHER, "%s: request to create name %s for "DLID4"\n",
2973 obd->obd_name, rec->ur_tgt, OLID4(rec->ur_id1));
2975 /* first, lookup the target */
2976 child_lockh.cookie = 0;
2978 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id2, parent_lockh,
2979 &de_tgtdir, LCK_PW, MDS_INODELOCK_UPDATE,
2980 &update_mode, rec->ur_tgt, rec->ur_tgtlen,
2981 &child_lockh, &de_new, LCK_EX,
2982 MDS_INODELOCK_LOOKUP);
2989 LASSERT(de_tgtdir->d_inode);
2992 rc = mds_add_local_dentry(rec, offset, req, rec->ur_id1,
2993 de_tgtdir, de_new, 0);
2998 if (cleanup_phase == 1) {
3000 if (parent_lockh[1].cookie != 0)
3001 ldlm_lock_decref(&parent_lockh[1], update_mode);
3003 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
3004 if (child_lockh.cookie != 0)
3005 ldlm_lock_decref(&child_lockh, LCK_EX);
3010 req->rq_status = rc;
3014 static int mds_reint_rename_to_remote(struct mds_update_record *rec, int offset,
3015 struct ptlrpc_request *req)
3017 struct obd_device *obd = req->rq_export->exp_obd;
3018 struct ptlrpc_request *req2 = NULL;
3019 struct dentry *de_srcdir = NULL;
3020 struct dentry *de_old = NULL;
3021 struct mds_obd *mds = mds_req2mds(req);
3022 struct lustre_handle parent_lockh[2] = {{0}, {0}};
3023 struct lustre_handle child_lockh = {0};
3024 struct mdc_op_data opdata;
3025 int update_mode, rc = 0;
3028 CDEBUG(D_OTHER, "%s: move name %s onto another mds #%lu\n",
3029 obd->obd_name, rec->ur_name, (unsigned long)id_group(rec->ur_id2));
3030 memset(&opdata, 0, sizeof(opdata));
3032 child_lockh.cookie = 0;
3033 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, parent_lockh,
3034 &de_srcdir, LCK_PW, MDS_INODELOCK_UPDATE,
3035 &update_mode, rec->ur_name,
3036 rec->ur_namelen, &child_lockh, &de_old,
3037 LCK_EX, MDS_INODELOCK_LOOKUP);
3040 LASSERT(de_srcdir->d_inode);
3044 * we already know the target should be created on another MDS so, we
3045 * have to request that MDS to do it.
3048 /* prepare source id */
3049 if (de_old->d_flags & DCACHE_CROSS_REF) {
3050 LASSERT(de_old->d_inode == NULL);
3051 CDEBUG(D_OTHER, "request to move remote name\n");
3052 mds_pack_dentry2id(obd, &opdata.id1, de_old, 1);
3053 } else if (de_old->d_inode == NULL) {
3054 /* oh, source doesn't exist */
3055 GOTO(cleanup, rc = -ENOENT);
3057 struct lustre_id sid;
3058 struct inode *inode = de_old->d_inode;
3060 LASSERT(inode != NULL);
3061 CDEBUG(D_OTHER, "request to move local name\n");
3062 id_ino(&opdata.id1) = inode->i_ino;
3063 id_group(&opdata.id1) = mds->mds_num;
3064 id_gen(&opdata.id1) = inode->i_generation;
3066 down(&inode->i_sem);
3067 rc = mds_read_inode_sid(obd, inode, &sid);
3070 CERROR("Can't read inode self id, "
3071 "inode %lu, rc = %d\n",
3076 id_fid(&opdata.id1) = id_fid(&sid);
3079 opdata.id2 = *rec->ur_id2;
3080 rc = md_rename(mds->mds_lmv_exp, &opdata, NULL, 0,
3081 rec->ur_tgt, rec->ur_tgtlen - 1, &req2);
3086 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3092 ptlrpc_req_finished(req2);
3095 if (parent_lockh[1].cookie != 0)
3096 ldlm_lock_decref(&parent_lockh[1], update_mode);
3098 ldlm_lock_decref(&parent_lockh[0], LCK_PW);
3099 if (child_lockh.cookie != 0)
3100 ldlm_lock_decref(&child_lockh, LCK_EX);
3105 req->rq_status = rc;
3109 static int mds_reint_rename(struct mds_update_record *rec, int offset,
3110 struct ptlrpc_request *req, struct lustre_handle *lockh)
3112 struct obd_device *obd = req->rq_export->exp_obd;
3113 struct dentry *de_srcdir = NULL;
3114 struct dentry *de_tgtdir = NULL;
3115 struct dentry *de_old = NULL;
3116 struct dentry *de_new = NULL;
3117 struct inode *old_inode = NULL, *new_inode = NULL;
3118 struct mds_obd *mds = mds_req2mds(req);
3119 struct lustre_handle dlm_handles[7] = {{0},{0},{0},{0},{0},{0},{0}};
3120 struct mds_body *body = NULL;
3121 struct llog_create_locks *lcl = NULL;
3122 struct lov_mds_md *lmm = NULL;
3123 int rc = 0, cleanup_phase = 0;
3124 void *handle = NULL;
3127 LASSERT(offset == 1);
3129 DEBUG_REQ(D_INODE, req, "parent "DLID4" %s to "DLID4" %s",
3130 OLID4(rec->ur_id1), rec->ur_name, OLID4(rec->ur_id2),
3133 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
3135 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
3136 DEBUG_REQ(D_HA, req, "rename replay\n");
3137 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
3138 lustre_msg_buf(req->rq_reqmsg, offset + 3, 0),
3139 req->rq_repmsg->buflens[2]);
3142 MD_COUNTER_INCREMENT(obd, rename);
3144 if (rec->ur_namelen == 1) {
3145 rc = mds_reint_rename_create_name(rec, offset, req);
3149 /* check if new name should be located on remote target. */
3150 if (id_group(rec->ur_id2) != mds->mds_num) {
3151 rc = mds_reint_rename_to_remote(rec, offset, req);
3155 rc = mds_get_parents_children_locked(obd, mds, rec->ur_id1, &de_srcdir,
3156 rec->ur_id2, &de_tgtdir, LCK_PW,
3157 rec->ur_name, rec->ur_namelen,
3158 &de_old, rec->ur_tgt,
3159 rec->ur_tgtlen, &de_new,
3160 dlm_handles, LCK_EX);
3164 cleanup_phase = 1; /* parent(s), children, locks */
3165 old_inode = de_old->d_inode;
3166 new_inode = de_new->d_inode;
3168 /* sanity check for src inode */
3169 if (de_old->d_flags & DCACHE_CROSS_REF) {
3170 LASSERT(de_old->d_inode == NULL);
3173 * in the case of cross-ref dir, we can perform this check only
3174 * if child and parent lie on the same mds. This is because
3175 * otherwise they can have the same inode numbers.
3177 if (de_old->d_mdsnum == mds->mds_num) {
3178 if (de_old->d_inum == de_srcdir->d_inode->i_ino ||
3179 de_old->d_inum == de_tgtdir->d_inode->i_ino)
3180 GOTO(cleanup, rc = -EINVAL);
3183 LASSERT(de_old->d_inode != NULL);
3184 if (de_old->d_inode->i_ino == de_srcdir->d_inode->i_ino ||
3185 de_old->d_inode->i_ino == de_tgtdir->d_inode->i_ino)
3186 GOTO(cleanup, rc = -EINVAL);
3189 /* sanity check for dest inode */
3190 if (de_new->d_flags & DCACHE_CROSS_REF) {
3191 LASSERT(new_inode == NULL);
3193 /* the same check about target dentry. */
3194 if (de_new->d_mdsnum == mds->mds_num) {
3195 if (de_new->d_inum == de_srcdir->d_inode->i_ino ||
3196 de_new->d_inum == de_tgtdir->d_inode->i_ino)
3197 GOTO(cleanup, rc = -EINVAL);
3201 * regular files usualy do not have ->rename() implemented. But
3202 * we handle only this case when @de_new is cross-ref entry,
3203 * because in other cases it will be handled by vfs_rename().
3205 if (de_old->d_inode && (!de_old->d_inode->i_op ||
3206 !de_old->d_inode->i_op->rename))
3207 GOTO(cleanup, rc = -EPERM);
3210 (new_inode->i_ino == de_srcdir->d_inode->i_ino ||
3211 new_inode->i_ino == de_tgtdir->d_inode->i_ino))
3212 GOTO(cleanup, rc = -EINVAL);
3216 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
3218 * check if we are moving old entry into its child. 2.6 does not check
3219 * for this in vfs_rename() anymore.
3221 if (is_subdir(de_new, de_old))
3222 GOTO(cleanup, rc = -EINVAL);
3225 /* check if inodes point to each other. */
3226 if (!(de_old->d_flags & DCACHE_CROSS_REF) &&
3227 !(de_new->d_flags & DCACHE_CROSS_REF) &&
3228 old_inode == new_inode)
3229 GOTO(cleanup, rc = 0);
3232 * if we are about to remove the target at first, pass the EA of that
3233 * inode to client to perform and cleanup on OST.
3235 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
3236 LASSERT(body != NULL);
3238 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
3240 DOWN_READ_I_ALLOC_SEM(new_inode);
3242 cleanup_phase = 2; /* up(&new_inode->i_sem) when finished */
3244 if (new_inode && ((S_ISDIR(new_inode->i_mode) &&
3245 new_inode->i_nlink == 2) ||
3246 new_inode->i_nlink == 1)) {
3247 if (mds_orphan_open_count(new_inode) > 0) {
3248 /* need to lock pending_dir before transaction */
3249 down(&mds->mds_pending_dir->d_inode->i_sem);
3250 cleanup_phase = 3; /* up(&pending_dir->i_sem) */
3251 } else if (S_ISREG(new_inode->i_mode)) {
3252 mds_pack_inode2body(obd, body, new_inode, 0);
3253 mds_pack_md(obd, req->rq_repmsg, 1, body,
3254 new_inode, MDS_PACK_MD_LOCK);
3258 OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_RENAME_WRITE,
3259 de_srcdir->d_inode->i_sb);
3261 if (de_old->d_flags & DCACHE_CROSS_REF) {
3262 struct lustre_id old_id;
3264 mds_pack_dentry2id(obd, &old_id, de_old, 1);
3266 rc = mds_add_local_dentry(rec, offset, req, &old_id,
3267 de_tgtdir, de_new, 1);
3271 rc = mds_del_local_dentry(rec, offset, req, de_srcdir,
3276 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
3277 handle = fsfilt_start_log(obd, de_tgtdir->d_inode, FSFILT_OP_RENAME,
3278 NULL, le32_to_cpu(lmm->lmm_stripe_count));
3281 GOTO(cleanup, rc = PTR_ERR(handle));
3284 de_old->d_fsdata = req;
3285 de_new->d_fsdata = req;
3286 rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
3289 if (rc == 0 && new_inode != NULL && new_inode->i_nlink == 0) {
3290 if (mds_orphan_open_count(new_inode) > 0)
3291 rc = mds_orphan_add_link(rec, obd, de_new);
3294 GOTO(cleanup, rc = 0);
3296 if (!S_ISREG(new_inode->i_mode))
3299 if (!(body->valid & OBD_MD_FLEASIZE)) {
3300 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
3301 OBD_MD_FLATIME | OBD_MD_FLMTIME);
3302 } else if (mds_log_op_unlink(obd, new_inode,
3303 lustre_msg_buf(req->rq_repmsg,1,0),
3304 req->rq_repmsg->buflens[1],
3305 lustre_msg_buf(req->rq_repmsg,2,0),
3306 req->rq_repmsg->buflens[2],
3308 body->valid |= OBD_MD_FLCOOKIE;
3314 rc = mds_finish_transno(mds, (de_tgtdir ? de_tgtdir->d_inode : NULL),
3315 handle, req, rc, 0);
3317 switch (cleanup_phase) {
3319 up(&mds->mds_pending_dir->d_inode->i_sem);
3322 UP_READ_I_ALLOC_SEM(new_inode);
3325 if (dlm_handles[5].cookie != 0)
3326 ldlm_lock_decref(&(dlm_handles[5]), LCK_PW);
3327 if (dlm_handles[6].cookie != 0)
3328 ldlm_lock_decref(&(dlm_handles[6]), LCK_PW);
3331 ptlrpc_save_llog_lock(req, lcl);
3334 if (dlm_handles[3].cookie != 0)
3335 ldlm_lock_decref(&(dlm_handles[3]), LCK_EX);
3336 ldlm_lock_decref(&(dlm_handles[2]), LCK_EX);
3337 ldlm_lock_decref(&(dlm_handles[1]), LCK_PW);
3338 ldlm_lock_decref(&(dlm_handles[0]), LCK_PW);
3340 if (dlm_handles[3].cookie != 0)
3341 ptlrpc_save_lock(req,&(dlm_handles[3]), LCK_EX);
3342 ptlrpc_save_lock(req, &(dlm_handles[2]), LCK_EX);
3343 ptlrpc_save_lock(req, &(dlm_handles[1]), LCK_PW);
3344 ptlrpc_save_lock(req, &(dlm_handles[0]), LCK_PW);
3353 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3356 req->rq_status = rc;
3360 typedef int (*mds_reinter)(struct mds_update_record *, int offset,
3361 struct ptlrpc_request *, struct lustre_handle *);
3363 static mds_reinter reinters[REINT_MAX + 1] = {
3364 [REINT_SETATTR] mds_reint_setattr,
3365 [REINT_CREATE] mds_reint_create,
3366 [REINT_LINK] mds_reint_link,
3367 [REINT_UNLINK] mds_reint_unlink,
3368 [REINT_RENAME] mds_reint_rename,
3369 [REINT_OPEN] mds_open
3372 int mds_reint_rec(struct mds_update_record *rec, int offset,
3373 struct ptlrpc_request *req, struct lustre_handle *lockh)
3375 struct obd_device *obd = req->rq_export->exp_obd;
3376 struct lvfs_run_ctxt saved;
3379 /* checked by unpacker */
3380 LASSERT(rec->ur_opcode <= REINT_MAX &&
3381 reinters[rec->ur_opcode] != NULL);
3383 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);
3384 rc = reinters[rec->ur_opcode] (rec, offset, req, lockh);
3385 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &rec->ur_uc);